<%BANNER%>

A Lane-Changing Model for Urban Arterial Streets

Permanent Link: http://ufdc.ufl.edu/UFE0024835/00001

Material Information

Title: A Lane-Changing Model for Urban Arterial Streets
Physical Description: 1 online resource (237 p.)
Language: english
Creator: Sun, Jian
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: corsim, focus, gap, lane, microscopic, traffic, urban
Civil and Coastal Engineering -- Dissertations, Academic -- UF
Genre: Civil Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: As one of the most fundamental components in microscopic traffic simulation, lane-changing affects the distribution of vehicles across lanes and contributes to traffic movements. In recent years, the topic of lane-changing has become of increased importance in traffic engineering and safety research. Previous lane-changing models divided the behavior as either mandatory (MLC) or discretionary (DLC) based on the purpose of the maneuver. Generally, MLC occurs when drivers have to change lanes in order to keep the right route. DLC refers to cases in which drivers change lanes to acquire driving benefit, such as overtaking slow vehicles, bypassing a heavy vehicle, avoiding the traffic toward an off-ramp, and so on. It is well accepted that driver characteristics (such as level of aggressiveness, alertness) have large impact on various aspects of both maneuvers (MLC and DLC), such as the level of acceptance on a particular DLC, minimum or maximum acceleration/deceleration adopted, etc. However, the existing models have not incorporated the driver characteristics with much detail. This thesis contributes to the development of lane-changing models for urban arterials in microscopic traffic simulation. It enhances existing models and develops new ones as appropriate. In this research, the effect of driver characteristics was incorporated in modeling both the acceptance of various DLC reasons and the gap acceptance procedure within lane-changing maneuvers. To accomplish this, a focus group study was first carried out to capture behavior differences among drivers. Next, an in-vehicle field data collection was performed to investigate the effect of driver type on specific MLC and DLC scenarios, and collected microscopic data from the corresponding lane-changing maneuvers. With the field collected values, a comprehensive model was developed to handle the probability of changing lanes under each proposed DLC reason and the gap acceptance procedures. The lane-changing probability for each DLC scenario was modeled as a function of corresponding important factors (obtained from focus group) and driver types. In gap acceptance modeling, the hand-shaking negotiation concept (from the TCP/IP protocols in computer network communications) was introduced to describe the vehicle interactions during lane-changing maneuvers under congested traffic flow. The proposed lane-changing model was developed and implemented in a microscopic traffic simulator, CORSIM. Traffic data were collected along a congested arterial in the City of Gainesville, FL, and used for model calibration and validation purposes. Simulation capabilities of the newly developed model were compared against the original lane-changing model in CORSIM. The results indicate that the new model better replicates the observed traffic under different levels of congestion.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Jian Sun.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Elefteriadou, Ageliki L.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-08-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0024835:00001

Permanent Link: http://ufdc.ufl.edu/UFE0024835/00001

Material Information

Title: A Lane-Changing Model for Urban Arterial Streets
Physical Description: 1 online resource (237 p.)
Language: english
Creator: Sun, Jian
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2009

Subjects

Subjects / Keywords: corsim, focus, gap, lane, microscopic, traffic, urban
Civil and Coastal Engineering -- Dissertations, Academic -- UF
Genre: Civil Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: As one of the most fundamental components in microscopic traffic simulation, lane-changing affects the distribution of vehicles across lanes and contributes to traffic movements. In recent years, the topic of lane-changing has become of increased importance in traffic engineering and safety research. Previous lane-changing models divided the behavior as either mandatory (MLC) or discretionary (DLC) based on the purpose of the maneuver. Generally, MLC occurs when drivers have to change lanes in order to keep the right route. DLC refers to cases in which drivers change lanes to acquire driving benefit, such as overtaking slow vehicles, bypassing a heavy vehicle, avoiding the traffic toward an off-ramp, and so on. It is well accepted that driver characteristics (such as level of aggressiveness, alertness) have large impact on various aspects of both maneuvers (MLC and DLC), such as the level of acceptance on a particular DLC, minimum or maximum acceleration/deceleration adopted, etc. However, the existing models have not incorporated the driver characteristics with much detail. This thesis contributes to the development of lane-changing models for urban arterials in microscopic traffic simulation. It enhances existing models and develops new ones as appropriate. In this research, the effect of driver characteristics was incorporated in modeling both the acceptance of various DLC reasons and the gap acceptance procedure within lane-changing maneuvers. To accomplish this, a focus group study was first carried out to capture behavior differences among drivers. Next, an in-vehicle field data collection was performed to investigate the effect of driver type on specific MLC and DLC scenarios, and collected microscopic data from the corresponding lane-changing maneuvers. With the field collected values, a comprehensive model was developed to handle the probability of changing lanes under each proposed DLC reason and the gap acceptance procedures. The lane-changing probability for each DLC scenario was modeled as a function of corresponding important factors (obtained from focus group) and driver types. In gap acceptance modeling, the hand-shaking negotiation concept (from the TCP/IP protocols in computer network communications) was introduced to describe the vehicle interactions during lane-changing maneuvers under congested traffic flow. The proposed lane-changing model was developed and implemented in a microscopic traffic simulator, CORSIM. Traffic data were collected along a congested arterial in the City of Gainesville, FL, and used for model calibration and validation purposes. Simulation capabilities of the newly developed model were compared against the original lane-changing model in CORSIM. The results indicate that the new model better replicates the observed traffic under different levels of congestion.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Jian Sun.
Thesis: Thesis (Ph.D.)--University of Florida, 2009.
Local: Adviser: Elefteriadou, Ageliki L.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-08-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2009
System ID: UFE0024835:00001


This item has the following downloads:


Full Text

PAGE 1

1 A LANE-CHANGING MODEL FOR URBAN ARTERI AL STREETS By DANIEL(JIAN) SUN A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2009

PAGE 2

2 2009 Daniel(Jian) Sun

PAGE 3

3 To my Mom

PAGE 4

4 ACKNOWLEDGMENTS I would like to take this oppor tunity to thank Professor L ily Elefteriadou for being a wonderful advisor, both technical and otherwise, a constant s ource of encouragem ent since my arrival at UF. Her innovative ideas provided the foundation of this research and her encouragement, meticulousness and pursuit for perf ection greatly enhanced its quality. It has been an honor and life-changing experience working with her. I thank the other members of my doctoral committee: Mr. William M. Sampson, Professors S cott S. Washburn, Orit Shechtman and Si va Srinivasan, for thei r critical review of different aspects of this research and assistance. Mr. Sampson had been my supervisor during the period I was working in Mc Trans Center. I really a ppreciated the huge sup ports and assistances obtained. Dr. Siva has been a friend, philosophe r and guide from my first year at UF. His technical and practical insights in driver behavior modeling were i nvaluable for this research. Dr. Washburn is an excellent mentor and instructor, who provided numerous suggestions that led to the complete to this thesis. Dr. Shechtman also provided many helpful suggestions from different perspectives. I benefited a lot from her experiences in the focus group study and in-vehicle experiment. I thank TRC and CMS faculty, staff and student s, too many to list, who made this an enjoyable experience. Special thanks to my fellow students and staff at the Mc Trans Center for their friendship. The moments of joys and disappoint ments that we shared together will be one of the greatest treasures of my life. I was extremely fortunate to have an amazing group of Chinese friends who eased stressful times and extended their warmth for their advice re garding my career and research. Above all, I am grateful to my parents: XianZhi Sun and ZaiHua Zhang, and my sister for their endless encouragement and unconditional support for my intellectual pursuits.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........8 LIST OF FIGURES.......................................................................................................................11 ABSTRACT...................................................................................................................................13 CHAPTER 1 INTRODUCTION..................................................................................................................15 1.1 The Problem Statement................................................................................................. 15 1.2 Research Objectives...................................................................................................... 17 1.2.1 Evaluate the Impact of Driver Characteristics on the Lane-Changing Maneuver..........................................................................................................17 1.2.2 Develop a Probabilistic Model for Each of DLC Reasons............................... 18 1.2.3 Develop a Gap Acceptance Model for Different Lane-Changing Modes......... 18 1.3 Thesis Outline...............................................................................................................19 2 LITERATURE REVIEW.......................................................................................................20 2.1 Rule-Based Microscopic Lane-Changing Models........................................................ 20 2.1.1 Gipps Lane-Changing Model........................................................................... 21 2.1.2 Weis Heuristic Structured Lane-Changing Model...........................................22 2.1.3 Multi-Agent Lane-Changing Model Used in ARTEMiS.................................. 24 2.1.4 Other Rule-Based Lane-Changing Models....................................................... 30 2.1.5 Commercial Simulators with Rule-Based Lane-Changing Models.................. 32 2.2 Discrete Choice-Based (DCB) Mi croscopic Lane-Changing Models.......................... 35 2.2.1 Lane-Changing Model Used in MITSIM.......................................................... 35 2.2.2 Other Recent DCB Lane-Changing Models..................................................... 37 2.3 Other Microscopic La ne-Changing Models.................................................................. 45 2.4 Summary and Conclusions............................................................................................48 2.5 Recommendations......................................................................................................... 49 3 METHODOLOGY................................................................................................................. 51 3.1 Research Step 1 Focus Group Study and Information Categorization....................... 53 3.2 Research Step 2 In-Vehicle Field Data Collection and Results Analysis............... 56 3.3 Research Step 3 Lane-Changing Probability Model and Gap-Acceptance Model..... 61 3.3.1 Lane-Changing Probability Model....................................................................61 3.3.2 Gap Acceptance Model.....................................................................................63 3.4 Research Step 4 Model Implementation and Validation............................................ 65

PAGE 6

6 3.4.1 CORSIM Implementation................................................................................. 66 3.4.2 System Validation Procedures.......................................................................... 66 3.5 Summary and Conclusions............................................................................................69 4 FOCUS GROUP-BASED STUDIES.....................................................................................71 4.1 Focus Group Preparation and Implementation............................................................. 72 4.1.1 Preparation of Questions................................................................................... 72 4.1.2 Participant Recruitment and Prescreening........................................................ 80 4.1.3 Other Issues....................................................................................................... 83 4.2 Analysis of the Results.................................................................................................. 84 4.2.1 Driver Type Classification Scheme..................................................................84 4.2.2 Probability of Various Actions for Different Driver Types.............................. 91 4.2.3 Critical Factors for Each Lane-Changing Scenario..........................................93 4.3 Summary and Conclusions............................................................................................98 5 IN-VEHICLE EXPERIMENT AND ANALYSIS...........................................................100 5.1 In-Vehicle Experiment Preparation and Implementation........................................ 100 5.1.1 Participant Characteristics............................................................................... 101 5.1.2 Testing Route..................................................................................................104 5.1.3 Driving Test Procedure...................................................................................110 5.2 Data Reduction and Analysis...................................................................................... 112 5.2.1 Video Data Reduction..................................................................................... 113 5.2.2 Distributions of Selected Lane-Changing Variables.......................................117 5.2.3 Cluster Analysis for Driver Type Classification............................................. 125 5.2.3.1 Classification scheme I driver background-based scheme............ 125 5.2.3.2 Classification scheme II dr iver behavior based scheme..................128 5.2.3.3 Results comparison........................................................................... 132 5.3 Summary and Conclusions..........................................................................................135 6 MODEL DEVELOPMENT.................................................................................................. 137 6.1 Scenario-Based Lane-Changing Probability Model....................................................138 6.1.1 Dataset Overview............................................................................................ 138 6.1.2 Lane-Changing Probability Function Estimation............................................ 141 6.2 Gap Acceptance Model for Urban Arterials............................................................... 156 6.2.1 Gap Acceptance Characteristics...................................................................... 157 6.2.2 Notations and Modeling Framework.............................................................. 160 6.2.3 Decision Framework of the Subject Vehicle S1..............................................162 6.2.4 Decision Framework of the Lag Vehicle T2...................................................164 6.2.4.1 Competitive behavior........................................................................ 166 6.2.4.2 Cooperative behavior........................................................................168 6.3 Summary and Conclusions..........................................................................................169

PAGE 7

7 7 MODEL IMPLEMENTATION AND VALIDATION........................................................ 172 7.1 Datasets....................................................................................................................... 172 7.2 CORSIM Implementation and Calibration................................................................. 174 7.2.1 CORSIM Implementation............................................................................... 175 7.2.2 Aggregate Calibration..................................................................................... 182 7.3 Systematic Validation................................................................................................. 183 7.3.1 Comparison of the Lane-Based Travel Time.................................................. 184 7.3.2 Comparison of the Lane Distribution.............................................................. 188 7.3.3 Comparison of the Vehicle-Based Cumu lative Number of Lane Changes.....190 7.3.4 Sensitivity Analysis......................................................................................... 191 7.4 Summary and Conclusions..........................................................................................195 8 CONCLUSIONS.................................................................................................................. 198 8.1 Research Summary......................................................................................................198 8.2 Contributions............................................................................................................... 201 8.3 Directions for Future Research................................................................................... 204 APPENDIX A MEMORANDUM OF THE SUBMIT MATERIALS CHECKLIST .................................. 206 B UFIRB PROTOCOL FORM................................................................................................ 208 C INFORMED CONSENT FORM FOCUS GROUP STUDY ............................................ 210 D INFORMED CONSENT FORM IN-VEHICLE EXPERIMENT ................................. 212 E ADVERTISEMENT FLYER FOR THE PARTICIPANTS RECRUITMENT................... 214 F PRESCREENING QUESTIONAIRE FOR PARTICIPANTS SELECTION..................... 215 G PARTICIPANTS DRIVING BACKGROU ND SURVEY QUESTIONAIRE .................... 217 H FOCUS GROUP MODERATIING SCRIPTS..................................................................... 218 I K-MEAN ALGORITHM USED TO OBTAIN THE CENTROIDS ................................... 227 LIST OF REFERENCES.............................................................................................................228 BIOGRAPHICAL SKETCH.......................................................................................................237

PAGE 8

8 LIST OF TABLES Table page 3-1 Reason-based lane-changing in formation table (for Reason n) ......................................... 62 3-2 Reason-based probability functions es timated from the in-vehicle data........................ 63 4-1 Focus group categories and questions................................................................................ 74 4-2 Form for documenting the level of likelihood for DLC reasons....................................... 76 4-3 Factors and the respect ive importance for a given lane-changing situation...................... 80 4-4 Personal background information of the focus group participants.................................... 82 4-5 Driver-based likelihood of execu ting a discretionary lane change.................................... 85 4-6 Driver type categorization by the characte ristics demonstrated in verbal expression....... 90 4-7 Consistency between the clustering result and driver type demonstrated......................... 90 4-8 Lane-changing likelihood level for each driver group (L1-L4).........................................92 5-1 Overview of the partic ipants characteristics for the in-vehicle experiment................ 101 5-2 Personal background information of the In-Vehicle experiment participants............. 102 5-3 Detailed route information for the In-Vehicle data collection experiment (Newberry Road route).................................................................................................... 108 5-4 Detailed route information for the InVehicle data collection experiment (Waldo Road route).......................................................................................................................109 5-5 In-Vehicle experiment notes for subject ID: 05-11...................................................... 111 5-6 Driver-based number of maneuvers collected in the In-Vehicle experiment............... 113 5-7 Data collected from comp leted lane changes (ID = 0511).............................................. 116 5-8 Data collected from attempted but unsuccessful lane changes (ID = 0511).................... 117 5-9 Data collected from poten tial lane changes (ID = 0511).................................................117 5-10 Statistics of variables rela ted to completed lane changes................................................ 118 5-11 Statistics of variables rela ted to attempted lane changes................................................. 121 5-12 Statistics of variables rela ted to potential lane changes................................................... 122

PAGE 9

9 5-13 Driver-Based likelihood of executin g a DLC (In-Vehicle experim ent).......................126 5-14 Drivers FAI interpolated fr om selected field behaviors................................................. 130 5-15 Consistency between the background-based and behavior-based classifications............133 5-16 Statistical distribution of dr ivers by the two classifications............................................ 134 6-1 Classification of driver groups for the in-vehicle experiment (based on FAI)............ 139 6-2 Number of lane changes collected for each scenario....................................................... 140 6-3 Number of LCs for different driver types during Stopped-Bus scenario..................... 142 6-4 Estimated coefficients for the factors in Stopped Bus scenario................................... 143 6-5 Number of LCs for different driver types during Vehicle Merge scenario.................. 144 6-6 Estimated coefficients for the f actors in Vehicle Merge scenario................................ 145 6-7 Number of LCs for different driver types during Slow Vehicle scenario.................... 146 6-8 Estimated coefficients for the factors in Slow Vehicle scenario.................................. 147 6-9 Number of LCs for diff erent driver types during Q ueue Advantage scenario............. 148 6-10 Estimated coefficients for the f actors in Queue Advantage scenario........................... 148 6-11 Number of LCs for different driver types during Heavy Vehicle scenario.................. 150 6-12 Estimated coefficients for the factors in Heavy Vehicle scenario................................ 150 6-13 Number of LCs for different driver types during Tailgating scenario......................... 151 6-14 Estimated coefficients for the factors in Tailgating scenario....................................... 152 6-15 Number of LCs for different driv er types during Pavement scenario.......................... 153 6-16 Estimated coefficients for the factors in Pavement scenario........................................ 154 6-16 Scenario-based utility functions )( LCV estimated from the in-vehicle data.............. 155 6-18 Number of LCs for different driver types during Back Turning scenario.................... 155 6-19 Number of LCs for different driver types during Pedestrian/Scooter scenario............ 155 6-20 Number of LCs for different driver types during Erratic Drivers scenario.................. 155 6-21 Number of completed/attempted LC s based on modes and driver types......................... 158

PAGE 10

10 6-22 Observed spacing gap characteristics for the com pleted/attempted lane changes........... 159 6-23 Observed accelerations/decelerati ons within vehicle interactions................................... 159 7-1 Summary of the Newberry Road video data.................................................................... 174 7-2 Initial and calibrated values of the pa rameters in CORSIM lane-changing model......... 176 7-3 Travel time and number of lane-changing measurements............................................... 176 7-4 Detailed strategy on distributing CORSIM drivers into the clustering groups................ 181 7-5 Initial and calibrated values of the pa rameters in the new lane-changing model............ 183 7-6 Comparison of simulation travel time between new and CORSIM models (2 Test).................................................................................................................................186 7-7 Comparison of simulation travel time be tween new and CORSIM models (T Test).. 187 7-8 Goodness-of-fit statistics for lane-based travel speeds, vehicle counts and number of lane changes.....................................................................................................................193

PAGE 11

11 LIST OF FIGURES Figure page 1-1 Four steps involved in a general lane-changing m aneuver................................................ 16 2-1 Flowchart of the heuristic lane-changing maneuver model............................................... 23 2-2 Flowchart of the lane-changing process in ARTEMiS...................................................... 25 2-3 The lane-changing model struct ure from Ahmeds dissertation........................................ 39 2-4 Combined lane-changing model in NGSIM...................................................................... 44 2-5 Hierarchical framework for a general lane-changing model............................................. 50 3-1 Proposed framework......................................................................................................... .52 3-2 The HTRD 400 system and other equipmen t (i.e. GPS, DCs) in the Honda Pilot............ 59 3-3 Image-based vehicle distance estimation........................................................................... 60 3-4 Possible interaction scheme within cooperative model..................................................... 65 3-5 Procedures included in the systematic validation.............................................................. 68 4-1 Typical lane-changing scenar ios occurred on urban streets.............................................. 77 4-2 Results for the clustering with different number of cluster............................................... 88 5-1 Proposed route for the field data collection (Newberry Road route)............................... 105 5-2 Proposed route for the field da ta collection (Waldo Road route).................................... 106 5-3 Vehicles involved in a lane-cha nging maneuver and related variables........................... 118 5-4 Distributions of lane-changing variables for the completed maneuvers.......................... 120 5-5 Distributions of lane-changing variables for the attempted maneuvers........................... 123 5-6 Distributions of lane-changing variables for the potential maneuvers............................ 124 5-7 Clustering analysis result s based on driver background..................................................128 5-8 Clustering analysis result s based on driver behavior....................................................... 132 6-1 Modeling framework for choices of pl an and action in lane-changing behavior............ 137 6-2 Initial scenario and notations adopted for the gap acceptance model.............................. 160

PAGE 12

12 6-3 Framework of the lane-changing algorithm..................................................................... 161 6-4 Gap length on the target lane for different lane-changing modes.................................... 162 6-5 Effective components included in the lag spacing headway........................................... 165 6-6 Decision tree for modeling competitive/cooperative behavior........................................ 166 6-7 Competition scenario in the competitive/cooperative lane changes................................ 166 6-8 Cooperation scenario in the comp etitive/cooperative lane changes................................ 169 7-1 The Newberry Road segment for data collection............................................................ 173 7-2 Sketch of the segment selected for va lidation data collecti on (not to scale)................... 173 7-3 Volume data from video reduction taken on May 3rd 2005 PM peak period..................175 7-5 Implementation of the lane-chang ing decision procedure in CORSIM........................... 178 7-6 Schematic representation of a lane change in CORSIM.................................................. 180 7-7 CORSIM driver classification based directly on the FAI values..................................... 180 7-8 CORSIM driver classification based directly on the sampling percentages.................... 181 7-9 Volume data from video reduction taken on April 30th, 2005 PM peak period.............. 184 7-10 Comparison of the lane-based average travel time.......................................................... 185 7-11 Comparison of the lane distribution................................................................................. 189 7-12 Comparison of the cumulative number of lane changes by vehicles............................... 191

PAGE 13

13 Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy A LANE-CHANGING MODEL FOR URBAN ARTERIAL STREETS By Daniel(Jian) Sun August 2009 Chair: Lily Elefteriadou Major: Civil Engineering As one of the most fundamental components in microscopic traffic simulation, lanechanging affects the distribution of vehicles across lanes and contri butes to traffic movements. In recent years, the topic of lane-changing has become of increased importance in traffic engineering and safety research. Previous lane-changing models divided the behavior as either mandatory (MLC) or discretionary (DLC) based on the purpose of the maneuver. Generally, MLC occurs when drivers have to change lanes in order to keep the right route. DLC refers to cases in which drivers change lanes to acquire dr iving benefit, such as overtaking slow vehicles, bypassing a heavy vehicle, avoiding th e traffic toward an off-ramp, and so on. It is well accepted that driver characteristics (such as level of aggressiveness, alertness) have large impact on various aspects of both maneuvers (MLC and DLC), such as the level of acceptance on a particular DLC, minimum or maximum accelerati on/deceleration adopted, etc. However, the existing models have not incorporated the driver characteristics with much detail. This thesis contributes to the development of lane-changing models for urban arterials in microscopic traffic simulation. It enhances existing models and develops new ones as appropriate. In this research, th e effect of driver characteris tics was incorporated in modeling both the acceptance of various DLC reasons a nd the gap acceptance pr ocedure within lane-

PAGE 14

14 changing maneuvers. To accomplish this, a focus group study was first carried out to capture behavior differences among drivers. Next, an in-vehicl e field data collection was performed to investigate the effect of driver type on specific MLC and DLC scenarios, and collected microscopic data from the corresponding lane -changing maneuvers. With the field collected values, a comprehensive model was developed to handle the probability of changing lanes under each proposed DLC reason and the gap acceptance procedures. The lane-changing probability for each DLC scenario was modeled as a functi on of corresponding important factors (obtained from focus group) and driver types. In gap accep tance modeling, the hand-shaking negotiation concept (from the TCP/IP protocols in comput er network communications) was introduced to describe the vehicle interactions during lane-c hanging maneuvers under conge sted traffic flow. The proposed lane-changing model was deve loped and implemented in a microscopic traffic simulator, CORSIM. Traffic data were coll ected along a congested arte rial in the City of Gainesville, FL, and used for model calibration and validation purposes. Simulation capabilities of the newly developed model were compared against the original lane-changing model in CORSIM. The results indicate that the new model better replicates the observed traffic under different levels of congestion.

PAGE 15

15 CHAPTER 1 INTRODUCTION 1.1 The Problem Statement During last several decades, a large amount of work has been done to formulate models of traffic flow and build traffic simulation applicati ons (Chandler et al., 1958; Herman and Rothery, 1969; Gipps, 1981, 1986; Barcelo et al., 1996; Yang and Koutsopoulos 1996; Zhang et al., 1998; Hidas, 2002, 2005; Liu et al., 2006). Car-followi ng and lane-changing are two most fundamental components in microscopic traffic simulation. Car-following models deal with the time and space relationships of two consecutive vehicles in the same lane, and control the motion of the lag car (Pipes, 1953; Newell, 1961; Gazis et al., 1961). Lane-changi ng affects the distribution of vehicles across lanes (Rorbech, 1976; Brackst one et al., 1998). Compared to car-following models, in which the behavior of the lead vehicle is relatively unaffected by the lag one, the lanechanging process depends on many parameters and hence it is more complex. Generally, a lane-changing process is modeled as a sequence of four decision-making steps, as shown in Figure 1-1. Step 1 considers whether a lane-changi ng is necessary and whether the potential lane-changing reasons are accepted by th e subject driver. Next, the target lane is determined for the lane-changing maneuver in Step 2. Step 3 checks the lead and lag gaps in the target lane, so that an appropriate lane change is chosen or rejected accordingly. In the last step, an acceleration or deceleration is ad opted to move the subject vehicle to the target lane. Each of these steps is formulated with the corresponding fiel d-collected or simulated data from interested transportation facilities, respectively. Many previous studies have focused on la ne-changing behavior along freeways (Ahmed 1996; Laval and Daganzo 2006), in which reasons invoking lane-changing are typically to gain speed advantage. Other research investigates the lane-changing behavior on freeway on-ramp

PAGE 16

16 merging area (Kita, 1999; Kita et al., 2002; Choudhury, 2005). For this case, drivers would recognize the necessity of performing a mandatory lane change as they arrive at the merging point. Figure 1-1. Four steps involved in a general lane-changing maneuver Despite the great significance, la ne-changing behaviors, especially those occurring in urban arterials, have not been studied as extensively as car-following behavior. Limited research has been reported regarding lane changes on urba n arterials, where the possible lane-changing instances are numerous (Hidas, 2005; Ben-Akiva et al., 2006). Only a few researchers have tried to address important issues of the congested conditions in urban st reets lane-changing maneuvers (Gipps, 1986; Yang and Koutsopoulos, 1996; Wei et al., 2000; Hidas, 2002). Even these models have ignored some complex factors within the lane-changing maneuver, such as the high level of interactions among vehicles and the driver behavior variability involved. Consequently, research documenting drivers thinking process, as well as support for the assumptions used in existing models, is scarce. One major reason for this is the scarcity of reliable data (Brackstone and McDonald, 1996; Hidas and Wagner, 2004). Data required to develop lane-changing models include the position, speed, acceleration, and length of a subject

PAGE 17

17 vehicle and the vehicles ahead of and behind the subject vehicle in the current lane as well as in the adjacent lanes. In addition, site-specific fact ors, such as speed limit of the segment and road geometry, also affect lane-chang ing behaviors. Data collected fr om cross-sectional detectors are not sufficient to report lane-changing procedures. In the recent ten years, with wide use of video devices in urban traffic areas and the existence of video tracking applications, traffic engineers can collect high quality vehicle trajectory data (Hoogendoorn et al., 2003), which are used to obtain detailed lane-changing maneuvers. In this thesis, a comprehensive framework fo r modeling drivers la ne-changing behavior on urban arterials is presented. Emphases are placed on the lane-changing reasons (Step 1 in Figure 1-1) and gap acceptance (Step 3 in Figure 1-1), in which any possible conditions may affect the drivers final decision. One majo r objective of this thesis is to model the drivers lane-changing behavior under congested traffic in microscopic perspective. Speci al attention is placed on the effects of driver characteristi cs on the lane-changing maneuvers. 1.2 Research Objectives The objective of this thesis is to develop a comprehensive model for drivers lane-changing behavior on urban arterials. The model in corporated both non-conge sted and congested conditions with special attenti on given to the impact of driv er characteristics on lane-changing behavior. More specifically, the three sub-objectives of this research are: 1.2.1 Evaluate the Impact of Driver Characte ristics on Urban Lane-Changing Maneuvers As one category of the ubiquitous and important factors in lane-chang ing behavior, driver characteristics (level of aggressiveness, aler tness, etc) are captured by two well-designed experiments, focus group study and in-vehicle data collecti on, conducted in this thesis research. Factors obtained from the focus groups include both driver behavioral parameters and environmental parameters that affect lane change s. An effective classification is then developed

PAGE 18

18 to categorize participating drivers into differe nt types based on the quantitative results and qualitative verbal expression c onclusions from the focus group st udy. Thus, various drivers can be invited to drive an instrumented vehicle, so that field data are collected for testing and validating the effect of driver type on specifi c lane-changing actions. 1.2.2 Develop a Probabilistic Model for Each of DLC Reasons Most existing models assume the decision of change lanes under each DLC scenario is deterministic (Gipps, 1986; Wei et al., 2000; Hidas, 2002, 2005), which means that the DLC reasons are always accepted when the given conditi ons are satisfied. However, the real situation is probabilistic and stochastic, since the decision depends on many interdependent factors. With the driver behavior information and quantitative data collected under each lane-changing scenario, a probabilistic model can be developed for each of the invoking DLC scenarios, in which the probability of changing lanes can be formulated as a function of surrounding traffic states and driver characteristics. The modeli ng coefficients can be estimated from the invehicle field lane-changing data, and then be calibrated with the additional source of field datasets. 1.2.3 Develop a Gap Acceptance Model for Different Lane-Changing Modes Drivers tend to behave differently and accep t different gap criteria under urban traffic conditions. In this thesis, three types of lane-changing modes: free, cooperative/competitive and forced, are defined, so that different gap accep tance procedures can be developed for each. The hand-shaking negotiation c oncept (Stevens, 1990, 1998) are adopted to describe communication between vehicles during lane -changing maneuvers under congested traffic flow, through which interactions among the mergi ng vehicle and the lag vehicles on the target lane are captured. In this framework, each vehicl e is modeled as an intelligent agent: a reactive, autonomous, internally-motivated entity that inhabits a dynamic, not fully predictable traffic

PAGE 19

19 environment (Weiss, 1999). Each agent can send a lane-changing request signal (turning signal) to one or multiple agent(s), and then the receive r(s) evaluate the request and respond accordingly. A detailed procedure of intera ction is modeled based on the fi eld observations, and the studies are described in subsequent chapters of this thesis. The components described above were developed and integrated into a comprehensive lane-changing model. As one of the highlights, dr iver characteristics aff ect not only the decision of lane-changing reasons but also the gap accepta nce procedure, which correspond closely to the Step 1 and Step 3 in Figure 1-1. The other two steps of the lane-changing maneuver, Step 2 and Step 4, are not within the emphases of this research. 1.3 Thesis Outline The remainder of this thesis is organized in se ven chapters. In Chapter 2, a literature review on existing microscopic lane-changing models is presented. Chapter 3 discusses the methodology for developing the lane-changing model in this thesis. Chapter 4 and Chapter 5 present the detailed procedures of the two experiments, fo cus group study and in-ve hicle data collection, for obtaining the lane-changing related behavior data, along with the an alysis of results. Development procedures for the scenario-based lane-changing probability model and the gap acceptance model are presented in Chapter 6. In Chapter 7, the two components are implemented and integrated into the lane-changing model within a microscopic traffic simulator, CORSIM. Various comparisons of the newly developed model against CORSIM original model are provided, in terms of goodness-of-fit of model estimation and simulation capabilities. Finally, conclusions and directions for further research are summarized in Chapter 8.

PAGE 20

20 CHAPTER 2 LITERATURE REVIEW In this chapter, a lite rature review of lane -changing models is presen ted. Traditional rulebased lane-changing models, which are widely used in the existing micro-simulators, are discussed in Section 2.1. Secti on 2.2 reviews discrete choice based lane-changing models. Such a method is used in the Next Generation Simula tion (NGSIM) research (B en-Akiva et al., 2006). Section 2.3 presents other extensively research ed microscopic lane-cha nging models. Findings from the literature review are summarized in Section 2.4, followed by recommendations on a general lane-changing framework for micro-simula tion provided at the end of the chapter. 2.1 Rule-Based Microscopic Lane-Changing Models Since the early 1980s, the subj ect of lane-changing has received increased attention because of the technological progress on reliab le data collection (Brackstone and McDonald, 1999). Several realistic rule-based lane-changing algorithms have been developed. These algorithms have the ability to replicate drivers actions at the microscopi c level, and therefore can be incorporated to model lane-changing behavi or in micro-simulators. Additionally, the rulebased models can be calibrated using basic as sumptions about driver behavior, and can be verified using field data. Conse quently, they have been widely used in commercial and research packages, such as MULTSIM (Gipps and Wilson, 1980; Gipps, 1986), MITSIM (Yang and Koutsopoulos, 1996; Yang et al., 2000), AimSUN2 (Barcelo et al., 1996) and CORSIM (Halati et al., 1997). In this section, three types of rule-based lane-c hanging algorithms (Gipps model, Weis heuristic model and Hidas multi-agent mode l) are discussed in de tail. Other rule-based lane-changing models are reviewed briefly since no many details ar e available. Applications of rule-based lane-changing mode l in several well known commer cial micro-simulators are presented at the end of the section.

PAGE 21

21 2.1.1 Gipps Lane-Changing Model Gipps model (Gipps, 1986) is the first rule-based model that is well documented, and widely adopted in commercially available mode ls. By connecting lane-changing decisions to urban driving situations, Gipps mo del incorporates the most important factors, such as existence of safety gap, location of permanent obstructions intent of turning movement, presence of heavy vehicles and speed advantage. Based on the judgme nt on these criteria, the subject drivers decide whether to move to the target lane or not. The lane-changing reasons provided in Gipps model are as follows: Avoiding permanent obstructions; Avoiding the presence of special pur pose lanes such as transit lanes; Turning at the downstream intersection; Avoiding a heavy vehicles influence; and Gaining speed advantage. Gipps car-following formula (Gipps, 1981) wa s adopted to calculate suitable gaps between the subject vehicle and th e lead/lag vehicle(s), as well as the deceleration/acceleration required. The formula assumes that the driver of the following vehicle selects his/her speed to ensure he/she can bring the vehicle to a safe stop should the vehicle ahead come to a sudden stop. Thus, vehicle deceleration is used to evaluate the feasibility of changing lanes. The subject vehicle is assigned a sp ecial braking rate (nb ), from which a maximum deceleration for a given lane-changing maneuver can be obtained. If the d eceleration required for a lane change is not within the acceptance range, the lane change for th e subject vehicle is determined as not feasible. Gipps lane-changing model allows driver s to alter the brake rate parameter nb depending on the urgency of the lane-changing maneuver. The model equation is as follows: ]10/))((2[* nbVtxDbn nn n (2-1) where,

PAGE 22

22 nb is the special braking rate th at a maximum deceleration for a given lane-changing maneuver can be obtained, )( txDnn is the distance between the intended maneuver location and the current vehicle location, nV is the desired speed (free fl ow speed) of the driver, and nb is the most severe brakin g the driver would be willing to undertake. The lane-changing process in Gipps model can be summarized as a decision tree with a series of fixed conditions, wh erein situations that may be encountered on the road (urban arterials) were considered. Invoki ng of lane changing is a rule-b ased triggered event, and the final output is a binary choice (c hange/not change). The overall st ructure is flexible, and any new or special reasons for lane-cha nging can be added or replaced. However, this model does not consider the variability in indi vidual driver behavior, especi ally the different interaction strategies among the subject vehicle and th e surrounding vehicles under various traffic conditions. For example, under peak c ongested traffic, either the lag vehicle in the target lane has to consent to the lane change, or the subject vehicle has to force its way into the target lane. 2.1.2 Weis Heuristic Structu red Lane-Changing Model Based on videotaped observations over eigh t multi-lane urban streets in Kansas City, Missouri, Wei et al. (2000) deri ved a heuristic structure for rule s of a lane-changing model. In addition to the mandatory lane-changing (MLC) and discretionary lane-changing (DLC), a new type, named preemptive lane-changing (PLC), wa s defined when a vehicle explores lanechanging to the desirable lane if acceptable gaps are available. The intention is not for an immediate turn, but to proceed through the next intersection and make a turn at the following intersection. It was found that for this long-term lane-chang ing motivation, drivers accept different gap criteria (Wei et al., 2000). The heuristic structure of the model is presented in Figure 2-1. First, a lane change is categorized as MLC, PLC or DLC according to the maneuver

PAGE 23

23 intention and present location of the subject vehicle. Then three types of headways: T_Ld (to the lead vehicle in the target lane), T_Lg (to the lag vehicle in the target lane) and H_T (to the current head vehicle) are compared to the corresponding thresholds (1MLC 1PLC 3DLC which were estimated from field data). If all thre e headways are larger than the given thresholds, the lane-changing is acceptable and is complete d within a given time interval decided by the lane-changing type and the travel speed. Otherwise, the subject vehicle has to wait until the next time step and re-examine the lane-changing type and a new lane-changing maneuver is then initiated accordingly. Figure 2-1. Flowchart of the heuristic lane-changing maneuver model (source: Wei et al., 2000) In this model, for the MLC, additional speed and headway adjustments (m seconds, which may be calibrated from field data) are included after the lane-changing co nditions are checked as

PAGE 24

24 acceptable. This is because compared to the DLC and PLC, MLC always takes place within more strict conditions with smaller thresholds, and consequently drivers need to adjust their acceleration or decelera tion accordingly. A lane-changing duration is predefined for all maneuvers based on the vehicle speed. If the speed is larger than 7 mph, the time is set as 2.3 2.5 seconds. Otherwise, it is between 3.0 and 7. 5 seconds, as a function of vehicle position, velocity and acceleration. This lane-changing model was proposed as an important component in a dynamic laneassignment on given urban street networks, with which a simulation was developed to represent travel behaviors at lane level. Field-observed trajectory data were used to estimate the thresholds for all types of lane changes. However, this model only addressed the gap acceptance portion of lane-changing modes, and did not consider the reasons for lane-changing. Similar to Gipps model, Weis model did not cons ider interactions and communicat ion among vehicles, and is not able to reflect real lane-changing behavior under congested conditions. 2.1.3 Multi-Agent Lane-Changing Model Used in ARTEMiS By analyzing the data collected from videorecording, Hidas (2002) found that most urban drivers had to force their way into the destin ation lane during congested conditions, which was not modeled effectively by previous lane-cha nging algorithms. In ARTE MiS (previously call SITRAS), Hidas (2002) adopted the autonomous agent technique to model drivers interactions involved in a more complex deci sion-making process, in which each vehicle was modeled as a driver-vehicle object (DVO) If a DVO perceives that another DVO intends to move into its lane, it may act as giving way, slowing down or not giving way, depending on road congestion conditions and individual driv er characteristics. This lane -changing decision process was presented in Figure 2-2. Similar to other rule-b ased models, the reasons for lane-changing were first evaluated, and the results were classified as Essential, Desirabl e or Unnecessary. The

PAGE 25

25 target lane was then selected according to the purpose of lane-cha nging. Different gap acceptance models were used for different la ne-changing modes. The following lane-changing reasons were adopted by Hidas (modifie d based on Gipps lane-changing model). Turning movement or the end-of-lane, Blockage or other obstructions, Transit or car pool lane, Speed advantage, and Queue advantage. Figure 2-2. Flowchart of the lane-changing process in ARTEMiS (source: Hidas, 2002) Two lane-changing modes were proposed according to the traffic conditions and the necessity of changing lanes: Normal lane change: In ARTEMiS, a normal lane change is considered when there is a gap of sufficient size in the target lane so that the s ubject vehicle can move in without forcing other vehicles in the target lane to slow down significantly. This can be expressed in two conditions as: a) Using the car-following model (Hidas, 1998), the deceleration (or acceleration) required for the subject vehicle to move behind the leader vehicle in target lane is acceptable, and

PAGE 26

26 b) Using the car-following model, the deceleration required for the lag vehicle in target lane to allow the subject vehicle to serve as its leader is acceptable. The car-following algorithm used in the model is provided as follows: )(*)()(1TtDTtxTtxn n n (2-2) where )( txi is the position of vehicle i at time t; T is the time interval that the following vehicle ( n ) attempts to reach a desired spacing; )( tDi is the desired spacing of vehicle i at time t; and is the driver judgment error parameter. This equation assumes that when approaching an d following a leader vehicle (n-1) at any time t, the driver of the following vehicle (n) attempts to adjust its acceleration so as to reach a desired spacing after a time lag of T seconds. The acceleration/decelera tion of the following vehicle can be calculated as follows: 1 2 2 1 2 1 25.0 5.0 ) ( 5.0 1 )( 5.0 n n nnnnn n nn n na TT T vxx TT vv TT T a (2-3) where iv is the speed of vehicle i; and n and n are the desired spacing constants of the follower vehicle n. Using Eq. (2-2) and Eq. (2-3), the deceleration required for the subject vehicle and the lag vehicle can be calculated. The two decelerati ons are then compared with the acceptable deceleration, which is calculated using a modifi ed format originally suggested by Gipps (1986): *)]*10/())((2[ LC n n nbVtxDb (2-4) where nb is the acceptable deceleration of vehicle n; D is the location of the intended turn or lane blockage; )( txn is the location of vehicle n at time t, nV is the desired (free) speed of vehicle n, LCb is the average deceleration a vehicle is willing to accept in lane changing, and

PAGE 27

27 is the driver aggressiveness parameter, calculat ed from the ratio of the subject driver type value to the average driver type value. Courtesy/forced lane change: The courtesy/forced lane-cha nging algorithm simulates the subject vehicle sending a courtesy signal to the subsequent vehicl es in the target lane. Starting from the first lag vehicle, the deceleration requi red to allow the subjec t vehicle to merge is calculated by the car-following model as described in Eq. (2-2). An acceptable deceleration nb is then calculated by Eq. (2-4). Once the new followe r is found, the new leader vehicle is the one right in front of the follower. By applying the car -following algorithm to the new leader vehicle, the subject vehicle and new lag vehicle, a gap of sufficient size will be created and the subject vehicle will move into the target lane. In another more recent paper, Hidas (2005) fu rther classified lane-changing maneuvers into three categories: free, cooperative and forced. The lead and lag gaps are us ed as the criteria of lane-changing feasibility checking. Based on the status of the th ree vehicles before changing lanes, the lead and follow gaps at the end of lane-changing were calculated as lg and fg. If both gaps are larger than the desired critical gaps (),(min, slead llvvgg and ),(min, lagsffvvgg), a free lane-changing is feasible. If this condition is not sati sfied and the lane-changing is essential for the subject vehicle, cooperative (courtesy) or forced lane-changing needs to be checked. The desired gaps min, lgand min, fgare the summation of a cons tant (minimum safe gap) and a value in direct proportion of speed difference, as follows. otherwise 0 if )(* ),(min min, leads leads lead slead lvvvvc gvvg (2-5) and otherwise 0 if )(* ),(min min, slag slag lag lagsfvvvvc gvvg (2-6)

PAGE 28

28 where ming is the minimum safe constant, leadc and lagc are the coefficients for the speed differences, and leadv lagvand sv are the travel speeds for lead, lag and subject vehicles, respectively. For the cooperative lane-changing, both the wi llingness of the lag driver and the feasibility of the maneuver need to be checked. A certain maximum speed decrease (vD ) from the lag vehicle is selected to indicate the willingness, which is a function of a vehicles aggressiveness parameter and the urgency of lane-changing. By setting the deceleration period as fvbDt / the lag gap at the end of deceleration can be ca lculated, which is the smallest gap between the subject vehicle and the lag vehicle after changing lanes. If this gap is larger than the minimum acceptable lag gap (min, ffgg ), a cooperative lane-changing is recognized as feasible. The forced lane-changing is similar to the coopera tive one, and differs only in that the maximum speed decrease (vD ) and deceleration fbare assumed by the subject vehicle as average values. Hidas (2005) validated the model using ve hicle trajectories from 73 lane-changing maneuvers in the Sydney CBD, Aust ralia. A total of four hours of video recording was collected from a road section where lane changing or merging maneuvers occurred. The tapes were first viewed and a number of lane-changing maneuvers were identified. Then, each maneuver was analyzed in detail, and the position and speed of each vehicle involved in the maneuver were identified at 0.2 s intervals using frame-by-frame analysis. The criteria parameters for gap acceptance (such as vD ming, leadc and lagc) were estimated from the video data and modified in the simulation according to individual drivers a ggressiveness. However, it is not clear how the driver aggressiveness was obtaine d from the data and related to these parameters. A simulation in ARTEMiS was run to test the cooperative an d forced lane-changing be havior for a freeway on-ramp situation with gradually increasing input flow rates. Th e speed and gap curves of both

PAGE 29

29 modes showed similar trends and maneuvers as the field-collected data. A weaving section was also simulated to determine the effect of th e lane-changing model on the relationship between average speed and flow rate. Results from multiple runs with and without the cooperative/forced modes showed that the full lane-changing model generated a speed-flow cu rve consistent with the expected shape, while the model without co operative/forced modes led to highly congested traffic at much lower flow rates. In the research, Hidas found that by only usi ng the trajectories from video data, the distinction between forced and cooperative lane changing may be ambiguous. It was concluded by his work that new empirical methods should be designed to collect lane changing related data. Several disadvantages relating to the lane-changing modeling in ARTEMiS were provided as follows: The given lane-changing reason set is incomple te. Some reasons, such as giving way to a merging vehicle or to a bus merging from a bus pull-off, or avoiding heavy vehicle influence, were not considered. Only the lag vehicle has the ability to initiate a cooperative lane-change. During the simulation, in each time interval all vehicles regardless of wh ether they are involved in a lane-changing maneuver are checked with respect to their intention to change lanes. In a situation where a free lane-changing is impossible, if the subject vehicle is checked first, a forced lane change is invoked. Otherwise, if the follower is checked first, it provides courtesy and adopts a cooperative mode. In reality, the interacti on between the subject vehicle and the lag vehicle includes two po ssibilities: cooperation or non-cooperation. The communication may last several seconds Referring to the steps in the TCP communication protocol (3-step handshaking negotiation), intera ctions between the subject vehicle and the follower can be interpreted as : the subject vehicle first sends a request signal to the lag vehicle. The lag vehicle eval uates the request and decides to decelerate accordingly. Then in the following time interv al, the subject vehicle re-evaluates the new gap and speed for the lane change. If the criteria are satisfied, a coopera tive lane change is executed. Otherwise, the slowing down of the follo wer vehicle continues. The critical gap values shown in Eq.s (2-5) and (2-6) are the summation of two components: the minimum safe constant gap mingand a value proportioned to the speed difference. However, in addition to the speed difference, the travel speed also affects the minimum acceptable gaps, and should be considered.

PAGE 30

30 2.1.4 Other Rule-Based Lane-Changing Models As mentioned at the beginning of this section, the rule-b ased lane-changing model is widely used in existing micro-simulators. In ad dition to the three models reviewed in detail, other models have also been proposed and studie d. For these models, no detailed implementation information is available, and consequently these algorithms cannot be duplicated. Thus, only briefly discussions are pr ovided in this section. Zhang et al. (1998) developed a multi-regime traffic simulation model (MRS), in which two types of lane changes, MLC and DLC, are de fined similarly to other models. The critical gaps of MLC are randomly distributed with a m ean estimated as a function of the remaining distance to the point where the lane change must be completed. Drivers in MLC situations may adjust their acceleration in order to be able to make the exis t gaps acceptable. The following cases are considered: No change in acceleration: The adjacent gap is acceptable. The subject needs to accelerate: Either the total length of the adjacent gap is sufficient but the lag gap is too small, or the total leng th of the adjacent gap is unacceptable but the gap ahead of the lead vehicle is acceptable. The subject needs to decelerate: Either the total length of the adjacent gap is sufficient but the lead gap is too small, or the total length of the adjacent gap is unacceptable but the gap after the lag vehicle is acceptable. The algorithm was implemented (Visual C++/MFC) using the input and output structure of CORSIM. The .trf files from CORS IM user graphical editor (TRAFE D) were used as inputs, and outputs similar to .tsd files were generated for TRAFVU animation. A post-processor was developed to provide additional MOEs. However, compared to CORSIM or other simulators, the simulation functions provided by MRS are very limited. During the model validation, only freeway data were used, with no real-world surface street data. Consequently, the model performance on urban streets, especially for congested tra ffic is difficult to assess.

PAGE 31

31 The micro-simulator DRACULA (Liu et al., 2006) developed in the University of Leeds, UK, integrated individual drivers day-to-day ro ute familiarity and route choice models with a traffic micro-simulation model of the car-fol lowing and lane-changing behaviors. The lanechanging model within the simulation firstly iden tifies the lane-changing desire according to a predefined set of rules. Once a lane-changing de sire is triggered, a gap-acceptance model is adopted to find the gaps in the target lane. A va riational critical gap is modeled to reflect the phenomenon of impatient drivers for whom the critical gap decreases with increasing waiting time. The stimulus required to induce the decrease of critical gap is modeled as the time spent searching for acceptable gaps. A minimum gap is used to set a lower boundary to the gapreduction formulation. The major functionality of DRACULA is to model individual trip makers decisions and the vehicle movements acr oss the network. The la ne-changing model is relatively simplified because the authors wanted to integrate driver familiarity and route choice into micro-simulation. Two disadvantages are found: first, not all major reasons for lane changing are considered. Second, similar to most of the previous algo rithms, the lane-changing model in DRACULA does not incorporate driver ch aracteristics, and no driver interactions are included. Hence, it does not replicate the re al-world urban traffic fully and accurately. Another microscopic traffic simulation and assignment model, INTEGRATION, considers the lane-changing desires as mandatory or discre tionary (Rakha and Zhang, 2003). To determine whether a DLC should be made or not, the perceived speeds in the current lane, the left adjacent lane and the right adjacent lane are compared by every second. Passenger cars have priorities to travel toward the middle lanes for roadways with three or more lanes. Trucks are biased toward using the shoulder lane. In situat ions where a trip destination imposes a constraint on vehicle movement, MLCs are performed to ensure that ve hicles maintain appropriate lanes (Prevedouros

PAGE 32

32 and Wang, 1999). Two boundaries were assigned upstream of the diverge point as softwall and hardwall. The hardwall, located closer to the diver point, indicates the location where subject vehicles are unable to proceed closer to the dive rge section on the original lane, and thus must abandon the lane. The softwall defines the location wher e the driver recognizes the need to change lanes. The mean locations of the soft wall/hardwall are at a distance of 100*n/10*n times the distance headway under jam density conditions (n is the minimum number of lane changes required to complete the maneuver). Because the two boundaries (softw all and hardwall) are used, the model is appropriate for modeling lane -changing behavior in w eaving sections. Similar boundaries may be brought into th e route-deciding MLCs IN urba n arterials, such as when changing lanes for a upcoming left/right turn. However, these two boundaries are not appropriate for most DLCs, during which the driver may ev en choose not to change lanes. The driver characteristics are more important in these situations. 2.1.5 Commercial Simulators with Ru le-Based Lane-Changing Models In CORSIM (FHWA, 1998) lane changes are cl assified as mandatory (MLC), discretionary (DLC) or random (RLC). The definitions of ML C and DLC are the same as in the previous models. RLC is performed by drivers for no appare nt reason, which may or may not result in an advantage for the vehicle over its current positio n. CORSIM assigns stochastically a certain percentage of drivers who perform such a random lane change (default value is 1 percent). For a vehicle performing either a RLC or DLC, it needs to stay in the lane for a given time period (the default value is 3 seconds). MLC is not subject to this and may be performed in any time step in response to the downstream geometrics In fact, the subject vehicle can change more than one lane in one time step in MLC. For any lane-changing maneuver in CORSIM (MLC, DLC or RLC), acceptable lead and lag gaps must be available in the target lane. Acceptance of the lead gap is modeled through the

PAGE 33

33 amount of the deceleration that is required by the subject vehicle to avoid collision with its leader in the target lane. The target leader is a ssumed to decelerate with the maximum possible deceleration, and the deceleration required by the subject vehicle in order to avoid collision is computed. This computed deceleration is compared to an acceptable dece leration which is called the acceptable lane changing risk. The lead gap is accepted if the required deceleration is smaller than the acceptable risk. A vehicl e with acceptable lead and traili ng gaps initiates a lane change into the target lane. Both the FRESIM and NETSIM adopt a similar lane-changing algorithm. The only difference lies in that the gaps in N ETSIM are measured in terms of time differences, and the gaps in FRESIM are a function of both ti me headways and speed differences between the subject vehicle and the lead and lag vehicles in the target la ne. One advantage of the lanechanging model in CORSIM is the flexibility of using user provided parameters. However, it does not consider the variability in gap acceptanc e behavior. The behavior is not modeled in a systematic manner, and all drivers are assumed to have identical gap acceptance behavior. Researchers in Transport Simulation System s (TSS) modified the Gipps lane-changing algorithm and incorporated it in to AIMSUN2 (Barcelo et al., 1996; Barcelo et al., 1998; TSS, 2004). Lane-changing in AIMSUN2 is modeled as a decision process evaluating the necessity of changing lanes, the desirability of changing lanes (lane-changing reasons), and the feasibility conditions for the lane change (the availability of gaps) depending on the location of the vehicle on the road network (Barcelo et al., 1996). Two braking values are calculated to decide whether a lane changing is possible. One is the braking im posed by the lead vehicle in the target lane to the subject vehicle, and the other is the braking imposed by the subject vehicle to the lag vehicle in the target lane. If both brak ing ratios are acceptable, the lane -changing is possible. A special on-ramp lane-changing model is designed to take into account whether a vehicle is stopped or

PAGE 34

34 not, whether it is at the beginni ng of the on-ramp queue or not providing it is stopped, and how long it has been waiting. Another vehicle para meter, maximum waiting time, determines how long a vehicle is willing to wait before getting impatient. After this time, the vehicle becomes more aggressive and will redu ce the acceptable gaps. Although it is stated that AIMSUN2 can model incidents, no detailed information is give n in any of the references on how the model deals with lane-changing under incident situations. The lane-changing model in VISSIM is composed of a complex set of rules, which depends much on the type of streets and other paramete rs (Fellendorf, 1994; PTV, 2004). For example, if a faster driver approaches a slower one on the sa me lane, it checks if it can improve the position by changing to a neighboring lane. The differe nce between freeways and urban arterials is considered significant in VISSIM. In urban streets, the next turni ng direction is one of the most important parameters for deciding the present la ne. Some other driver vehicle parameters are considered important in the VI SSIM lane-changing model, including : 1) technical description of a vehicle, 2) behavior of a dr iver, and 3) interaction between several drivers. The parameter minimum headway (front/rear) defines the minimum distance that must be available for a lanechanging in standstill conditions. In PARAMICS, two types of lane-changing are defined as overtaking lane-changing corresponding to the reason of speed advantage, and directional lane -changing corresponding to route choice reasons (Cameron and Duncan 1996; Quadstone, 2004). The minimal lanechanging gap is a combination of a user-defined value and individual driver type, and is provided in units of time. The lane-changing maneuver is completed successfully if a suitable gap exists continuously within a preset simulation interval required to complete the maneuver. The mean of this interval is four seconds, and the value in creases when the vehicle speed becomes lower. Two

PAGE 35

35 important link-related indices determine when a dr iver recognizes a required lane change. One is the distance that the most aware driver will s ee the oncoming required lane change, and the other is the distance that the least awar e drivers will see the required lane change. For the distances in between, drivers will proportionally see the re quired lane change. If the distance to the downstream junction is within a given user-defined value, the vehicle will cease to make any overtaking lane-changing decisions, and only cons ider the directional lane-changing until the turn at the next junction. This distance corres ponds to the real distan ce from the position of a directional signpost at the roadside to the junction. The drivers aggressiveness is taken into account in modeling this signposti ng lane-changing behavior. 2.2 Discrete Choice-Based (DCB) Microscopic Lane-Changing Models This section presents the literature review on the discrete-choi ce-based lane-changing models proposed by researchers in MIT and th e Next Generation Simulation (NGSIM) program. 2.2.1 Lane-Changing Model Used in MITSIM A microscopic traffic simulator, MITSIM (Y ang and Koutsopoulos, 1996), was designed to establish a laboratory environment for tes ting and evaluating new algorithms in ATMS (Advanced Transportation Management System) and ATIS ( Advanced Traveler Information System ). Similar to previous models, the lane-changing model in MITSIM was implemented to include three steps: (1) checking the necessity of lane-changing, (2) selecting the desired lane, and (3) checking whether gap distances are acceptable or not. Two types of lane changes, mandatory and discretionary, were defined. Fo r the mandatory one, the lane-changing maneuver starts at a distance nx from the downstream node (or incident, lane drop, red LUS) with probability given by the following equation: 1 ]/)-exp[( 0 0 2 n 2 0 n xx xx xx pn n n (2-7)

PAGE 36

36 where np is the probability that vehicle n starts a mandatory lane-changing maneuver, nx is the distance from downstream node or lane blockage, 0x is the critical distance, which may be associat ed to the position of a particular message sign (such as final exit warning), and n is a variable defined as Km2n10 n(nm is the number of lanes need to cross; K is the traffic density of the segment; 0 1 and 2 are parameters). When a MLC is invoked, the status is kept until the desired lane change has been completed, or the vehicle has moved into the do wnstream link. Two parameters related to DLC are the impatience factor and the speed indifference factor. These are used to decide whether the speed difference between the current lane and ta rget lane are large enough to invoke a lane change. The target lane choice is based on mu ltiple criteria including lane-changing regulations, drivers lane privilege, lane c ongestion, current signal state, prev ailing traffic conditions, drivers desired speed and lanes maximum speed. Once the ta rget lane is decided, the lead and lag gaps in the target lane are checked. For the DLC, the minimum acceptable gaps are given by the following equation: n ngg (2-8) where, ng is the minimum gap that driver n consider to be acceptable for a discretionary lane change, g is the average acceptable gap, and n is an error term. Both g and n are parameters provided as user input fo r both lead and lag gaps. For the MLC, the minimum acceptable gaps may decrease as the vehicle approaches the downstream node (same for incidents and lane drops), whic h is given by the following equation: *) ( min min max min min max min min max min max max xx g xxx xx xx ggg xx g gn n n n nn (2-9)

PAGE 37

37 where, ng is the minimum gap that driver n considers to be acceptable for a mandatory lane change, ming and maxgare the lower and upper bounds for gaps (lead and lag), nx is the vehicles current position, minx and maxx are the distances that define the range within which the critical gap varies from ming and maxg, and n is an error term. By these two equations, MITSIM models the difference of the gap acceptance between the two types of lane changes. That is, in a MLC drivers tend to accept sma ller gaps as they get closer to the last location where the lane cha nge has to take place. However, the reasons of changing lanes are not included in the model, which may have to be defined outside for simulation implementation. In addition, the model handles all issues from the field data, and does not capture driver characteristics. 2.2.2 Other Recent DCB Lane-Changing Models In Ahmeds dissertation (1999), a general framework that cap tures lane-changing behavior under both the MLC and DLC ( MLC) situations was developed. In this model, lane-changing is divided as a sequence of four st eps: 1) decision to consider a lane-changing, 2) target lane choice, 3) acceptance of gap conditions in the target lane, and 4) performing the lane-changing maneuver. A discrete choice concept was adopted to model the impact of the surrounding traffic environment and lane configuration as well as dr iver characteristics. The whole procedure is modeled as a decision tree, shown in Figure 2-3. Four layers, which correspond to the sequence of four steps, are included. The output from one layer is the input for the layer following. In the top level, a driver decides to respond to the MLC or DLC ( MLC), and the target lane is fixed in a MLC. Gaps in the target lane are checked a nd a lane change is i nvoked if the gaps are acceptable. In a DLC ( MLC), if a driver is not satisfied with the current lane, he/she will

PAGE 38

38 compare the driving conditions on the current lane with those of adjacent lanes and decide on a target lane. Similar to the MLC, the gaps in th e target lane are checked and a lane change is invoked if the gaps are acceptable. Otherwise, the driver will stay on the current lane. The probability of observing a change to the left lane is given by the following equation: ) |MLC(Pr *), MLC|DLC(Pr ) ,MLC DLC, |chosen laneleft (Pr ) ,MLC DLC, chosen, laneleft |accepable gap(Pr ) ,MLC DLC, chosen, laneleft accepable, gap |lanes (Pr ) |MLC(Pr ) MLC, |chosen laneleft (Pr ) MLC, chosen, laneleft |accepable gap(Pr ) MLC, chosen, laneleft accepable, gap |lanes change(Pr)|(Prn tn t n t n t n t n tn t n t n tntv v v v v change v v v v vL (2-10) where, nv is the individual specific random term, wh ich indicates the probability of decision to consider a lane change. Similarly, ) |Pr(ntnvJ for tnJ = R or C can be formulated (R: ri ght lane, and C: current lane). Finally, a likelihood function was formulated to estimate the related parameters: )()|(Pr*)|(Pr*)|(Pr11 nn N n T t nt nt ntdvvfvC vR vL Ln C tn R tn L tn (2-11) where L = change to the left lane R = change to the right lane C = continue in the current lane otherwise 0 tat time laneleft the tochangesn driver if 1L tn otherwise 0 tat time laneright the tochangesn driver if 1R tn otherwise 0 tat time lane changenot doesn driver if 1C tn )(nvf is the distribution of nv and N is the sample size.

PAGE 39

39 MLC Figure 2-3. The lane-changing model structure from Ahmeds disserta tion (source: Ahmed, 1999) In a numerical example, parameters of the model were estimated for a special simple case: merging from a freeway on-ramp. In this case, al l drivers initiate the change to the adjacent mainline as soon as they cross the merge point between the on-ramp and the freeway lane, and continue searching for acceptable gaps in the target lane. Although the given model provides a detailed framework for the lane-changing be havior research, no lane-changing reasons component is included, and the utili ty value for each candidate has to be calculated to obtain the lane-changing demand. Given the complexity of the lane-changing behavior on urban streets, it is difficult to acquire all necessa ry important factors to model all lane-changing situations. The author assumes that the existe nce (or non-existence) of an ML C situation is known. However, except for special cases, such as in the on-ramp merging example used, MLC situations can not be observed. Another weakness is that the mode l considers the lane-cha nging maneuver as a

PAGE 40

40 solitary decision, with no communication with other vehicles. This may be true for the highway traffic, where the sp eed is too high to allow comm unication among vehicles. For urban streets, in most situations, drivers inevit ably interact with each other for changing lanes, especially during congested traffic. To capture trade-offs between mandatory a nd discretionary consid erations, Toledo (2003) and Toledo et al. (2003) integrated the two into a single utility mode so that the awareness to the MLC situation is more realistically represented as a continuously increasi ng function rather than a binary choice. In this model, the lane-changing process consis ts of choice of target and gap acceptance decisions. Since traffic is a dynamic and uncertain environment, drivers in the algorithm need to re-evaluate and possibly modify their short-term goals and short-term plans as conditions change. Hence the lane-changing maneuver is modeled in a state-dependent stochastic manner. Under the discrete choice framework, the ut ilities of target lane and acceptable gaps are decided by vehicles surroundings, path plan and network knowledge and e xperiences. The target lane choice was formulated as a multinomial logi t model and the probabilities for each lane were given by the following equation: right} current, {left, ]|) exp[( ]|) exp[( )|(int TLi vv X vv X vTLPTLj nn TL j TL j TL jnt nn TL i TL i TL n i nt (2-12) where, TLXint is the vector of explanatory variables that affect the utility of lane i as a target lane to drive n at time t, TL i is the corresponding vector of parameters, nv is a driver/vehicle specific latent variable assumed to follow some distribution in the population, and TL i is the parameter of nv In this model, the choice of the target lane dictates the lane-changing direction. A gap acceptance model captures drivers choice by comparing the available gaps in the target lane

PAGE 41

41 with the critical gaps. The critical gaps ar e modeled as random variables with means being functions of explanatory variable s as in the following equation. left} {right, lag}, {lead, )ln(, d g v XGgd ntn gggd nt crgd nt (2-13) where crgd ntG, is the critical gap g in the direction of changed, measured in meters, gd ntX is a vector of explanatory variables, g is the corresponding vector of parameters, gd nt is a random term: gd nt~N(0, 2 gap), and g is the parameter of th e driver specific random term nv The gap acceptance was formulated as a multinomial probit model, which was affected by the spatial relations between the subject vehicle a nd the lead and lag vehicles in the adjacent lane. The lead/lag gap values were captured by variab les such as the subject relative speed and position with respect to the lead and lag vehicl es. The probability that gaps at time t are acceptable to driver n is given as: lag n lagdlag nt Tlag dlag nt lead n lead dlead nt Tlead dlead nt nnt crdlag nt dlag nt nnt crdlead nt dlead nt nnt crdlag nt dlag nt nnt crdlead nt dlead nt nnt nntv X G v X G vdG GPvdG GP vdGGPvdGGP vdgaplagacceptPvdgapleadacceptPP , )()ln( )()ln( ),|)ln()(ln(*),|)ln()(ln( ),| (*),| ( ),| (*),| ( (2-14) )( denotes the cumulative standa rd normal distribution. The model parameters were estimated usi ng data collected in a section of I-395 Southbound in Arlington, VA, which show that driver s lane selection is affected both by pathplan variables and surrounding tr affic environment. The critical gaps depend on the relative speeds with respect to the lead and lag vehicles. The implemen tation and validation results (in MITSIMLab) indicate the integrated model has better simulating ability for congestion build-up and dissipation. However, all estimation results of the model are based on freeway traffic.

PAGE 42

42 Additional factors, such as bus tr affic and pedestrian presence may need to be considered if the model is applied to urban street s. The model assumes that lane -changing maneuvers are carried out only when the acceptable gaps exist, which ma y not be the case in heav ily congested traffic. Forced merging and yielding need be considered for such conditions. To this point, this case study for a specific section of freew ay is not adequate, and more da tasets are required in order to identify geometry and other site-specific effects. With considering the large differences in the at tributes and utilities of the available lanes, Choudhury et al. (2004) and Toledo et al. (2005) developed a lane -changing model with explicit target lane choice, wherein the utility of each can didate lane was calculated from selected lane related variables and vehicle-specific attributes. This model adopts the same gap acceptance algorithm as Eq. (2-13) in the integrated model. The difference is, within this framework the candidate lanes include all eligible lanes on the ro ad. A driver may first change to a low utility lane in order to reach the target lane. The proposed lane-changing model was implemented in MITSIM lab (Choudhury et al., 2004). Validation sensor data and aggr egate trajectory data were collected from approximately 1.5 miles of highl y congested sections of I-80 in Emeryville and Berkeley, California. One set of aggregate data was used to calibrate the parameters included in the behavioral models in Eq. (2-13) (g). Then, a simulation was run with another input dataset from the same source. The performance of the target lane model was compared to the performance of the integrated lane-changing mode l with myopic change di rection (Toledo et al., 2003), i.e. the direction of the im mediate lane change is always the adjacent lane The validation process was based on the comparison of the simulated speeds and lane distributions. The results showed that the proposed lane-changing model provided significantly better prediction on the two indices. However, to apply the proposed mo del as a general lane-changing module to urban

PAGE 43

43 streets, sampling data are needed to estimate the coefficients in the discrete choice analysis. All existing parameter estimations for the model are based on freeway data, a nd special attributes (such as bus traffic and pedestri an) need to be incorporated to model the urban streets lanechanging behavior. Next, some attributes, such as trip planning, were chosen as important components of the model. Unfortunately, the effect s of how drivers adhere to the trip schedule were difficult to obtain. Finally, similar to the previous DCB models, the target lane model did not consider the forced and cooperative lane-changing behavior under congested traffic. Driver characteristics, which are very important in lane-changing maneuvers, were not considered in this model. Ben-Akiva et al. (2006) proposed a cooperative and forced merging model for MLC in the NGSIM report. The combined model is the first discrete choice model which considers forced or cooperative merging. It consists of three com ponents: normal lane-cha nging, cooperative lanechanging and forced lane-changing. A four-level decision-making process for the model is given in Figure 2-4.

PAGE 44

44 Figure 2-4. Combined lane-changing model in NGSIM (Source: BenAkiva et al., 2006) First, the target lead and lag gaps are comp ared to the normal lane-changing gap criteria, and a normal lane-changing will be executed, if both gaps are sufficient. Otherwise, the speed, acceleration and relative position of the lead and lag vehicles are checked. The anticipated lead and lag gaps are approximated by incorporating the courtesy from the lag vehicle. If the anticipated gaps are acc eptable, a cooperative lane-change under the perception of courtesy yielding from the lag vehicle will be initiated. If the anticipated gaps ar e still unacceptable, the driver will consider forced merging for MLC. Fo r a forced merging, the driver anticipates the deceleration that the lag driver would apply to accept him/her as a leader to avoid collision. The required deceleration of the lag vehicle is comp ared to the maximum deceleration a driver is willing to impose. If the antici pated deceleration is acceptable, a forced lane-change will be initiated. Otherwise, the merging vehicle must wait at the original lane with no lane change.

PAGE 45

45 The combined model considers MLC practically and the steps are rather clear. However, only the gap/deceleration acceptance is consider ed. The authors assume that the factors influencing drivers to change lane s are satisfied and the target lane has already been given. In the data validation, the freeway on-ramp trajectorie s were adopted so that both lane-changing reasons and target lane are pre-determined. Hence if being applied to the lane-changing maneuver in urban streets, additional modules are necessary. Even though the importance of drivers willingness and courtesy in the forced and cooperativ e lane-changing maneuvers was emphasized, the model did not mention how to obt ain and use the courtesy parameters, and how to determine to what extent the courtesy is provided from the lag vehicle. 2.3 Other Microscopic Lane-Changing Models In addition to the rule-based and the discrete-choice-based la ne-changing models introduced in the previous secti ons, other lane-changing models re lated to game theory, neural networks and kinematic waves have also been proposed. Kita (1999) and Kita et al. (2002) modeled the lane-changin g behavior at freeway on-ramp merging section based on game theory. The in teraction between two vehicles (the merging vehicle and the lag mainline vehicle) wa s modeled as a two-person non-zero-sum noncooperative game with complete information. The merging vehicle has two strategies: merge or pass, and the mainline vehicle can choose to give way or not. The strategy equilibrium (SE) is defined as a particular selecti on from mixed strategy choices of the players if and only if each player is using the best response to the strategy choices of the othe r players. No player can gain by unilaterally changing strategy. The data samples for validation were extracted from videotape observation, which recorded th e car movements in the sectio n by measuring the speed and headways between various vehicles and the end of merging section. The analysis data includes both actual choice, and theoretical choice pr obability based on th e observed surrounding

PAGE 46

46 conditions. The maximum likelihood method was used to estimate the parameter values for the explanatory variables in the payoff matrix. Comparison of the es timation results and field data indicate the proposed model ha s the capability to estimate th e probabilities of equilibrium selection of both players. The model can describe the traffic behavior in highway on-ramp merging sections, especially the interactions between the merging vehicle and the lag mainline vehicle during the lane-changing maneuver. However, it would be difficult to apply this model in urban streets. Because of the congested conditi ons in urban streets, a lag vehicle may choose slowing down instead of giving way to the subj ect vehicle. The purpose of this study was trying to understand the give way behavior by using activity survey data, and to provide a useful tool for describing interdependent dr iving behavior with interaction. Hence, this model was focused only on give way, and did not consider other behavior, such as slowing down under certain conditions. A simple assumption for the model is th at the driver will sel ect the action with the lower level of risk, where risk is expressed as the time to collision (TTC), and no minimal safe gap is considered. Also, the proposed model doe s not include the speed adjustment for the merging maneuver, which is an important part for changing lanes on urban arte rials, especially in congested traffic. Hunt and Lyons (1994) explored the application of neural ne tworks (NN) to model lanechanging decisions. A feed-for ward network trained using the back propagation learning algorithm was adopted. Input to the neural network at time t consisted of five sets of data, one for each consecutive time interval immediately proceeding t Each set includes the distance from the subject vehicle to the other 4-su rrounding vehicles (in the current lane or the target lane). Two other inputs are the current lane and speed of the subject vehi cle. The output of the neural network is the predicted speed and lane of th e subject vehicle for the next time interval t +1. This

PAGE 47

47 approach was implemented using video data collect ed from an urban street section. Comparison of the results from simulation and the real traffic data indicates the NN was able to correctly classify a high proportion of examples duri ng training for both simu lated and field data. However, the output predictions of speed and lane from the neural network are continuous variables, and a certain threshold must be impos ed to provide a specifi c lane output. Therefore when the output lane number is close to 1.5 (for a lane change from lane 1 to lane 2), each time interval could result in a different lane, suggest ing erratic driving behavior. A previous study of application of neural networks to robot carfollowing (Pomerl eau, 1992) also recognized that limitations in the training set could result in er ratic driving. In fact, microscopic lane-changing behavior is not deterministic but stochastic, wherein drivers characteristics and traffic environment should be considered. The NN in H unts model only takes the distances to the surrounding vehicles and the current speed as input, and no probabilistic conditions were included. Laval and Daganzo (2000) postulated freeway sections away from diverges, wherein the main incentive for drivers to change lanes is th e speed advantage. The lane-changing vehicle acts as a moving bottleneck on its destination lane until it accelerates to the prevailing speed, and additional lane changes may be triggered during this time period. Under such a situation, the freeway traffic is modeled as a set of interac ting streams linked by lane changes. The kinematic wave (KW) theory is introduced to treat the la ne-changing vehicles as a fluid that can accelerate instantaneously. By combining the KW theory w ith the accuracy of a mi croscopic model, slow vehicles are treated as moving bottlenecks in a KW stream. Under this framework, Laval and Daganzo (2006) further modeled each lane as a separate KW stream interrupted by lanechanging vehicles which allow no passing on the lane they occupy. Em pirical results were

PAGE 48

48 collected from two freeway secti ons, where obstructions with a ra nge of controlled speeds were introduced to simulate moving bottlenecks. The aut hors hold that by this model, the reduction in flow observed after the onset of congestion at freeway land-drops and the relationship between the speed of moving bottlenecks and their capacit ies can be explained. Additional simulations show that lane changes affect bottleneck behavior in ways that can be controlled to improve traffic flow (Tang et al., 2007). The emphasis of the KW model is the macroscopic relationships between flow, density and the ne t lane-changing rate, wherein a simplified linear flow-density relationship is adopted. For ideal freeway tra ffic, the simplification may not cause large discrepancy. However, when coming to the capr icious urban streets tr affic, especially for congested situations, the modeli ng ability is still in doubt. 2.4 Summary and Conclusions Lane-changing behavior has attracted more attention with th e development of microscopic traffic simulation tools. The ma neuver is usually classified as either mandatory (MLC) or discretionary (DLC). Each of these is further modeled in a sequence of three steps: 1) lanechanging necessity checking, 2) target lane c hoice and 3) gap acceptance decision. Based on a review of the existing lane-changing algorithms, the most popular ones are rule-based models and DCB models. Both of them may be implemented as a lane-changing module for the general micro-simulation. Rule-based algorithms model lane changes from the perspective of drivers. The reasons for lane changes are first enumerated and checked wh ether they apply. Then, the target lane is chosen from the adjacent lane(s), the parame ters for gap acceptance are retrieved from field/simulation data, and calcu lated by given formulas. Most of these parameters may be calibrated in the simulation. The DCB algorithms model driver behavior using logit or probit models, by which specific significant attributes are estimated. A drivers decision becomes a

PAGE 49

49 binary or multi-choice selection, an d utilities for all alternatives ar e calculated to get the output at each stage in the lane-changing process. Similar to the rule-based models, parameters for gap acceptance in the DCB models also need to be extracted from sampling data, and calibrated in the simulation. In addition to the rule-based and DCB models several other models were reviewed. The model based on game theory is largely limited to the merging-giveway behavior in freeway merging areas, and cannot be easily extended to other lane-changing ma neuvers. The neural network model tries to capture the relati onship between lane-c hanging maneuvers and driver/vehicle status (speed and distance to surr ounding vehicles). It may be applicable to the type of lane-changing maneuver for speed a dvantage purposes because of the distance consideration. However, the neural network algorithms do not incorporate the stochastic variability which is important in modeling comple x lane-changing behavior in urban streets. The basis for the kinematic-wave (KW) model is the assumption of a linear relationship between the flow, density and lane-changing ra te. This is a significant simplification of real traffic flow behavior, and is consequently difficult to be applied to the co mplex urban traffic. All these existing lane-changing models can not replicate lane changes on urban arterials accurately. Lane changes on arterial streets, especially under congested conditions, are characterized by several different invoking reasons, driver intera ctions, and a significant impact of driver characteristics. The rule-based model may be improve d by adding a driver interaction component to handle these characteristics. Additiona l effective field data are necessary, in order to model lane-changing invoking reasons and the impact of driver char acteristics correctly. 2.5 Recommendations Based on the literature review, a general la ne-changing framework for urban streets was developed and is summarized in Figure 2-5. The output of the first step is the determination of

PAGE 50

50 the lane-changing type (MLC, DLC or no change). For both the MLC and DLC, the target lane choice model and gap acceptance model need to be fo rmulated. If the gaps in the target lane are acceptable, a lane change is made. Otherwise, for the MLC, the vehicle must adjust its speed and position and attempt another MLC in the following time step. For a DLC, the vehicle may choose to give up the lane-changing attempt or re-evaluate the condition for a new DLC/MLC after the speed and position adjustment. Beginning of time step Step 1: Lane change decision Gap Acceptable? Y N Mandatory Lane Change Discretionary Lane Change No Change Left lane Right lane Step 2: Lane chosen Change to destination lane Gap Acceptable? Y N Change to destination lane Time Step N = N+1 Step 3: Gap condition checking Step 4: Lane change maneuver Position and Speed Adjustment Time Step N = N+1 Position and Speed Adjustment Condition 1 Condition 2 Otherwise Condition 1: Rout e choice, lane blockage, etc. Condition 2: Speed advantag e, queue advantage, etc. Left lane Right lane Check conditions on the other lane Y Change to destination lane N Give up lanechanging Figure 2-5. Hierarchical framewor k for a general lane-changing model Lane changing is a function of driving be havior. The framework given in Figure 2-5 provides a direction for further modeling procedures in this research. The difficulties lie in the uncertainty of environmental fa ctors and variations on driver and vehicle. New experiments should be designed to evaluate the impact of these factors, a nd consequently to improve the modeling capability of the algorithm.

PAGE 51

51 CHAPTER 3 METHODOLOGY Although driver behavior and characterist ics are important, they have not been incorpor ated into existing lane-changing models with much detail. The main reason is the diversity and uncertainty involved in human driv ing behavior. Research in artificial intelligence (AI) indicates that understanding human behavior is a very complex task, which makes modeling and automatic recognition of human activities difficult (Sim on, 1996). New methods should be designed to utilize detail ed driver behavior information in lane-changing modeling. In addition, more functional data need to be collected to assi st in the model development and implementation. This chapter presents a general methodology in developing a comprehensive lane-changing model for urban arterials, including the tasks at each research stage. Figure 3-1 presents the empirical data based research framework. The first step involves a focus group study, in which discussions among participants he lp to understand drivers con cerns under various lane-changing reasons and the corresponding behaviors. The re sults from the study ar e analyzed, with the expectation that would provide insights for developing a field data collection plan to be implemented in the step followed. In Step 2, part icipants are recruited to drive an instrumented vehicle, so that field values for specific importa nt factors identified in the focus groups can be collected. The field data are analyzed and categor ized by driver characteristics. Results collected from different driver groups were compared to decide the optimal classification scheme. Conclusions from this step ar e compared with the findings fr om focus group study to test and validate the effectiveness of the both experiments. In Step 3, the in-vehi cle field datasets are

PAGE 52

52 used to construct sub-models for modeling drivers decisions for particular DLC reasons and the gap acceptance procedure. These are eventually used to decide whether a lane change is necessary and the gaps in the target lane are accepta ble for changing lanes. Figure 3-1. Proposed framework

PAGE 53

53 Finally, in Step 4, the newly developed model is implemented and tested in a micro-simulator package, CORSIM. Aggregate statistics from vi deo data collected from Newberry Road in Gainesville, FL (Washburn and Kondyli, 2006), such as lane-based travel time, vehicle lane distributions and number of lane changes by vehi cles, are compared with the simulation results from the new model. Results using the existi ng lane-changing model in CORSIM are provided for comparison purposes. Calibration and validations, as well as a sensitivity analysis, are also conducted. Each of these four steps is further discussed in the remainder sections of this chapter, focusing on the motivation and the output of each step. Chapter 4 and Chapter 5 provide additional details for the efforts have been c onducted in focus group study and in-vehicle experiment, while the implementation and simu lation endeavors for model development and validation are discussed extens ively within the co rresponding Chapter 6 and Chapter 7. 3.1 Research Step 1 Focus Group Study and Information Categorization For a long time, field-based surveys (revealed preference/stated preferen ce) have been used to indicate drivers preferences for various tran sportation scenario s (Crane, 1996; Le vinson et al., 2004; Duan, 2006). For these surveys, however, the results may not be accurate either because the participants are not entirely truthful or they are rushed, an d have not had the opportunity to fully comprehend each question. During studying on self-disclosure, Jourard (1964) found that individuals decided to reveal based on their perceptions of th e other persons, and concluded subjects tended to disclose more about themse lves to people who resemble them. This is the theoretical foundation for focus group study, which has emerged as a form of qualitative research

PAGE 54

54 in recent decades. The difference between focu s groups and traditional survey is that focus groups involve a number of people at the same time, and discussions are highly encouraged (Stewart et al., 2007). What sepa rates focus groups from other interviewing methods is the fact that they allow group interaction, thereby providing greater insight into why certain beliefs and opinions are held (Loukopoulos et al., 2004). Focus groups encourage more critical thinking on the part of participants as a direct result of interacting with other participants and the moderator. Each of them may question, chal lenge or agree with the others beliefs and opinions. However, the purpose of focus groups is to listen and gather information without pre ssuring participants to vote or reach consensus. Group members can only influence each other by responding to ideas and comments of others. As the lane-changing behavior is affected by many interdepe ndent factors, and there may be large discrepancies among different types of drivers, focus groups are used in this research to obtain personal perceptions and attitudes regarding lane-changing maneuvers on urban streets. The objective of focus group studies in this research is to obtain the r easons and factors that affect the execution of lane cha nges. It is expected that this step results in providing lanechanging information from the general driving ex perience of each participant. The feedback and comments hopefully unveil a useful connection betw een driver type and la ne-changing factors. Each focus group meeting consists of the following three phases: Phase 1 A pre-selected list of lane-c hanging reasons is posed and th e participants are asked to prioritize them based on their driv ing experiences. The initial list is as follows (new reasons may be added during the discussion):

PAGE 55

55 R1.1Passing a stopped-bus at bus stop, R1.2Giving way to a merging vehicle, R1.3Gaining speed advantage by overtaking a slower moving vehicle, R1.4Gaining queue advantage, R1.5Avoiding a truck/heavy vehicle influence, R1.6Avoiding the pressure imposed by tailgating, and R1.7Attracted by a better pavement condition. Discrepancies may exist on the type and likelihood of DLCs for different of drivers. Five levels of likelihood (probabilities of invoking a lane ch ange for any given reason) were defined as: Level 1: Generally do not conduct (< 10%, weak), Level 2: Sometimes conduct but more likely do not (10% 40%), Level 3: Sometimes conduct, and sometimes do not (40% 60%), Level 4: More likely conduct (60% 90%), and Level 5: Generally conduct (> 90%, strong). In this step, answers from each participant are recorded for future lane-changing model development. The objective of this phase is to connect particular drivers to the acceptance for each of the pre-selected lane-c hanging reasons. The output is the level (i.e. probability) that a participant would change lane s for each of the identified re asons. Additional DLCs may be added by the participants at this stage. Phase 2 For each lane-changing reason, the participants are asked to provide factors that they consider when changing lanes. Th e list of DLCs was given as in Phase 1, and two MLCs are proposed as: R2.1Upcoming left/right turn at the imme diate/next downstream intersection, and R2.2Current lane is not available downstream (e.g. road incident, work zone or change in channelization of the current lane). Each of the participants discusses and desc ribes his/her behavior s under each lane-changing scenario. The significant factors are obtained, and would be used in developing guidelines for the in-vehicle field data collection. The objective of this phase is to connect each particular driver

PAGE 56

56 to the factors within each lane-changing reason. The output is a list of factors affecting the probability of changing lanes for each particular reason: Reason 1 (Factor11, Factor12, Factor1i), Reason 2 (Factor21, Factor22, Factor2j), Reason n (Factorn1, Factorn2, Factornk) Phase 3 The participants describe the possible interactions involved in lane changes under congested traffic. In this situation, cooperati ve and competitive strategies within lane-changing maneuvers are obtained. The objective of this step is to get interactions that may be involved in changing lanes, so that they could be modeled accurately. During the result analysis, drivers are tentativel y categorized based on their characteristics. Then the corresponding lane-changing behaviors is used to decide the optimal classification scheme, so that the driver characteristics for each driver type can be identified and incorporated in the further model development. The division on driver type is base d on specified level of driver aggressiveness, which was obtained through background survey. Outputs of the focus group study are used to develop gui delines for the in-vehicle fiel d data collection. A detailed experimental design and implementation for th e focus group study, along with the corresponding results analysis, are provided in Chapter 4. 3.2 Research Step 2 In-Vehicle Fiel d Data Collection and Results Analysis The focus group study provides the f actors that are important to drivers, as well as possible driver interactions during a lane-changing maneuver. The object ive of the in-vehicle data collection is to obtain field-meas ured values for the important f actors obtained in the focus group

PAGE 57

57 study. The field values collected in this step would be used to develop sub-models that can be incorporated into in a comprehensive lane-c hanging model. Results from the in-vehicle experiment are also used to test and va lidate conclusions from the focus group study. In this research, field data are needed to clarify following questions : 1) Whether a driver would accept a lane-changing reason? What is the probability? 2) Of the important factors proposed for a given lane-changing scenario, how do th ey affect the drivers decision? 3) For the different modes of lane changes, what gap crite ria are acceptable? and 4) How does the subject driver interact with the surrounding vehicles when a lane change occurs under congested traffic? As mentioned, the traditional cross-sectional detector data are not sufficient to describe lane-changing behavior, and no exis ting video trajectory data cover all lane -changing scenarios proposed in this research. Moreover, it is too costly and time-consuming to set up new video capturing facilities for the la ne-changing data collection. Consequently, a Honda Pilot instrumented vehicle was adopted to collect data related to the questions (1-4). In general, an instrumented vehicle is defined as a fully-operating, street legal vehicle that is equipped with flexible data acquisition systems to collect data such as speed, lane position, GPS, and driver eye movements, etc. (Chrysler et al., 2004). This technology allows the dr iving behavioral data to be safely collected while the subjects are in a natural state. More and more instrumented vehicles have been used in a va riety of traffic domains including driver behavior, road cataloging, and air quality studies, which help to gather more interesting da ta from drivers in a larger geographic area (Chrysler et al., 200 4). The instrumented vehicle technology is a useful tool for

PAGE 58

58 in-vehicle trajectory data colle ction, and has been used to co llect data regard ing car-following and lane-changing behaviors (Brackstone et al., 2002; Brackst one et al., 2009). During the driving test, each participant wa s accompanied by the researcher to drive on the pre-selected routes, where di fferent lane-changing scenarios may be invoked, and collect data related to the lane-chang ing maneuvers. Answers for questions 1) and 4) can be obtained directly from field observations, while for the rest two questions (2 and 3), special communications with the driver during the test were included to clar ify drivers decision process and actions. Notes are taken by the researcher to help the further data reduction procedure. In this research, a Honda SUV (Pilot 2006) equipped with a Honeyw ell Mobile Digital Data Recorder (HTRD400) system are used. Four wider coverage digital cameras (DCs) have been installed to capture video clips from different points of view, from which the necessary driving parameters and interactions with surr ounding vehicles can be inferred. As shown in Figure 3-2, DCs were installed to capture traffic in each dire ction (right, left, front and backward). Vehicle status data, including video and audio data, ar e recorded directly from the four HTRD cameras via standard CAT5 data /power cabling (Honeywell, 2005). From the GPS system connected, the HTRD400 can retrieve vehicle position, geographic direction, and speed data via a serial interface. Sp ecial software HTRD BusVie w connects the recorder to a PC/laptop. Video clips are first stored in the local hard disk, and can be transferred to a PC/laptop whenever needed. The player within the BusView can play video online or download clips to the local PC/laptop.

PAGE 59

59 Figure 3-2. The HTRD 400 system and other equipment (i.e. G PS, DCs) in the Honda Pilot The original video clips from DCs were taken at a 0.1 sec resolution. With the GPS system, the time-based location and speed data of the s ubject vehicle can be retr ieved. A new method has been designed to obtain car-follo wing and lane-changing data as follows. First, the video clips are decomposed into frame-by-frame images (0.5 sec each), so that the time-based relative position between vehicles can be observed. The targ et images are then selected for estimating the distance and speed of the surrounding vehicles based on the known distan ce in the camera view (such as lane width) and the dimensions of the instrumented vehicle (Figure 3-3). The computational steps include: 1) conducting some preliminary tests to estimate the camera constant c for each DC; 2) using the formula B B B BX x Y y Z c 0 to estimate the real distance B B Bx X cZ (all related variables are defined as in Figure 3-3). Consequently, the driver performance and the vehicle motions related to surrounding vehicles, such as reaction times, lane-changing gap acceptance, speed-distance relationships, speed perception, vehicle to vehicle communication links, etc. can be measured. With th is information, a dynamic time series record DC for monitoring left side traffic DC for monitoring right side traffic DC for monitoring traffic ahead

PAGE 60

60 of driver behavior relative to other vehicles can be obtained, which provides first-hand field data for the research on lane-changing behavior. (a) (b) c is the camera focus length constant, By is the y image coordinate of points 1B, 2B, OY is the camera height above ground level, Bx is the circumference dimension measured from the image (1B-2B), BX is the real value for circumference dimension, and BZ is the real distance from camera to target vehicle. Figure 3-3. Image-based vehicle distance estim ation (a) Geometry with horizontal camera (b) Measurements on the image Based on the field data collected from individual drivers, the fi eld values for different lanechanging scenarios and gap acceptance procedures ar e obtained. A result an alysis is adopted to categorize drivers based on their field lane-changing behaviors. The driver classification scheme is compared to the one obtained from the focus group study. If the two resu lts are not consistent with each other, it means either the categorization algorithm is not effective or the participants in any of the experiments do not behave themselves exactly. Iterative steps (experiments and classifications) have to be carried out until the two classification re sults are consistent.

PAGE 61

61 Outputs of the in-vehicle experiment ar e used to develop lane-changing model components to handle the probability of ch anging lanes under each scenario and the gap acceptance procedures. A detailed experimental design and implementation for the in-vehicle field data collection, as well as the corresponding results analys is, are provided in Chapter 5. 3.3 Research Step 3 Lane-Changing Prob ability Model and Gap-Acceptance Model The objective of this step is to develop th e sub-models for reason-based lane-changing probabilities and gap acceptance by using the re sults obtained from Step 1 and Step 2. 3.3.1 Lane-Changing Probability Model After successfully categorizing the lane-changing data by re asons, a set of reason-based lane-changing information can be attained. Table 3-1 presents an example of such information for a given reason n, in which the respective f actor values from all LC-related maneuvers are recorded (total number of maneuve rs is N). Each factor (Factorij) is associated with two indices: the first i is the reason category number (namely n in this case), and the second one j is the index within the reason n. The driver t ype information is determined by th e driver classification in Step 2. The number of factors listed would vary by sc enario, and would refer only to the important factors associated with that scenario. For a give n reason n, the value of each factor is obtained directly or indirectly from the in-vehicle field data collection. The probability of changing lanes under reason n can be formulated as a function of the respective attributes. The general format is as follows: P (lane-changing) = func(factor1, factor2, factorN, Driver type 1, Driver type 2, );

PAGE 62

62 For example, when it comes to the reason of Cu rrent lane is not ava ilable downstream because of road accident, the function and th e factor list may be proposed as: P (lane-changing) = func(factor1, factor2, factor3, Driver type 1, Driver type 2, ); where, P: is the probability to change lanes, factor1: the distance to the lane blockage location because of the road accident, factor2: level of congestion, factor3: relative speed, Driver type 1: the type of subject driver (1: belongs to driver type 1, 0: not), and Driver type 2: the type of subject driver (1: belongs to driver type 2, 0: not). Table 3-1. Reason-based lane-changi ng information tabl e (for Reason n) Factor S1 S2 S3 S4 S5 S6 S7 S8 S9 S10 S11 SN Factorn1 Factorn2 Factorn3 Factorn4 ... Driver type 1 Driver type 2 ... Note: For a give lane-changing reason n, each maneuver is used to generate the dataset including the values of each co rresponding important factor and th e type of the subject driver. Then a total of N datasets (samples) are used to develop the probability function of changing lanes under certain reason n. In addition to the set of reason-based lane-changing information illustrated in Table 3-1, a driver-reason relationship table is then constr ucted to store the proba bility functions and parameters for each reason (driver type i is one of the parameters). An example is provided in Table 3-2, in which for any given reason (j), a probability function of changing lanes (FLCj) is obtained. The functions and parameters are es timated from the reason-based lane-changing information table (Table 3-1). Fo r example, the function of FLCn and the corresponding coefficients for each factor (shaded cell in Table 3-2, for reason n) are estimated from

PAGE 63

63 information in Table 3-1 (N lane-changing mane uvers under reason n). Regression analysis is used to develop these functions Detailed analysis procedure is provided in the corresponding Chapter 6. Table 3-2. Reason-based probability functions estimated from the in-vehicle data Reasons Probability functions and parameters R1 FLC1: [10 DT1, DT2, ] R2 FLC2: [20 DT1, DT2, ] R3 FLC3: [30 DT1, DT2, ] Rn FLCn: [0 n DT1, DT2, ] In the model implementation, the function form ulations for the lane -changing probability model are stored as in Table 3-2. When the module is invoked within a micro-simulation, nondriver related parameters, such as average speed average travel time, av erage headway or queue length, are obtained in each simulation time step. Su ch information is used to calculate the values for the corresponding important at tributes (such as the factor2: level of congestion in the example above) within the reason. Based on these, including the specific driver type j (j [1, m]), the lane-changing probability for the given reason is calculated, so that th e lane-changing decision can be made. During simulation, all applicable re asons are checked for th e subject vehicle in a sequence. 3.3.2 Gap Acceptance Model It is well accepted that different modes, such as free, forced and cooperative ones, exist in lane-changing maneuvers (Hidas, 2005; Wang et al., 2005; Ben-akiva et al., 2006). The gap acceptance criteria may differ across the driver types and the lane-changing modes (Mahmassani and Sheffi, 1981). For example, maneuvers relate d to the free and forced lane changes are

PAGE 64

64 actions generally conducted in a comparably short time interval as soon as the decision has been made. The cooperative lane change includes inter actions with other sounding vehicles in several continuous seconds, and hence is more complex. One of the emphases in this research is to model the negotiation/competition procedures w ithin lane-changing maneuvers. The autonomous agent technique (Das et al., 1999) is adopted to model th e gap acceptance decision-making process. In the proposed algorithm, different lane-cha nging modes are decided as soon as a lanechanging reason is accepted. Each vehicle involve d in cooperative lane-changing is modeled as an intelligent agent. As presented in Figure 3-4, the merging agent (S1) looks for the lag agent (T2) and negotiates for a feasible lane ch ange. Two functions (MergingAgent and LagAgent) are developed to model the selfgovernment and communications between the two agents. Only the general framework is provided at this stage. The realistic details are retrieved from the focus group study and refined using th e field observations. The negotiation scenarios and lane-changing strategies of individual ve hicles are studied. The corresponding figures, flowchart and pseudo codes for the merging agen t and the lag agent are provided in the Chapter 6. For the free and forced lane changes, succe ssfully categorizing the results from focus groups and field observations help generate mode-bas ed critical gaps. The in itial critical gaps for these two modes can be obtained di rectly from the in-vehicle field data collection. The gaps include a deceleration index called the acceptable la ne-changing risk, which gets inflated as the need for changing lanes becomes urgent. Gap acc eptance in the cooperativ e lane-changing mode

PAGE 65

65 is modeled as a negotiation procedure, in which parameters, such as th e corresponding critical gaps and driver types, are used as inputs. The gap acceptance sub-model is going to be implemented as one of the key components in a comprehensive lane-changing model. Figure 3-4. Possible interaction scheme within cooperative model 3.4 Research Step 4 Model Implementation and Validation The objective of this step is to implement the proposed lane-changing model and validate it in CORSIM micro-simulator. First, each indivi dual component is implemented as a separate function within a lane-changing module, wh ich is invoked as a CORSIM RTE (run time extension) during the simulation. Next, field video data collected from arterials in Gainesville, FL, are used to calibrate CORSIM simulations. Two CORSIM simulation cases, with the newly developed lane-changing model or with the original CORSIM lane-changing model, are calibrated based the field data. Both calibrated models (with the new and original lanechanging algorithm) are simulated with the addi tional OD demands other than those calibration dataset, and the results are compared with the field measurements on multiple indices of the measures. Various statistical an alyses are conducted to evalua te the agreement between the

PAGE 66

66 results from the simulations and the field obser vations. The remaining of the section describes each step of the procedures. 3.4.1 Model Implementation Issues The integrated lane-changing model is implemented as a separate C++ DLL plug-in, which interfaces with CORSIM engine during the simu lation. Functions for each component are built and invoked as needed. In CORSIM simulation, the RT_PRE_NETSIM_VEHICLE message is sent just prior to calling the FORTRAN subroutine MOVE, which handles lane-changing, carfollowing, etc. to move all the vehicles for the current time step. The lane -changing plug-in is set up to respond to this message, and the function with in the plug-in is invoked to perform the lanechanging maneuver. By setting CO RSIM lane-changing timer to a value that would prevent the embedded lane-changing logic from being applied. The subroutine MOVE would still be called, but vehicles would not be allowed to make a lane change. The set up of the plug-in is the same as the configuration for general COSRIM RTEs. A detailed procedure was provided in CORSIM RTE Developers Guide (FHWA, 2006). 3.4.2 Model Validation Issues Field video data from the major arterials, su ch as Newberry Road and Archer Road, in Gainesville, Florida are collected The videos are first observed to identify the completed lanechanging maneuvers and the attempted but unsu ccessful lane-changes in volved. The position and speed of each vehicle involved in the maneuver can be obtaine d using frame-by-frame image analysis. Consequently, the maximum executing accel eration and deceleration in different types of lane changes can be inferred from the ti me dependent lead/lag gaps and speeds. Other aggregate parameters, such as flows and travel time can be obtained from the field video for

PAGE 67

67 calibration purposes. Additionally, various measures of performance, such as average lane-based speed, lane-based travel time, vehicle lane dist ributions and number of lane changes, for each individual arterial segment can be acquired for validation purposes. As presented in Figure 3-5, the test procedures are designed to validate the capability of the new model as follows. First, the arterial for data collection are si mulated with the field-measured OD demand in CORSIM. The simulation is calibrated by using the measurement from aggregate field data, such as average travel time and average speed. The process of calibration ai ms to adjust various parameters simultaneously, so that field observed traffic conditions can be accurately replicated. During this calibration, only the driver behavioral related parameters within CORSIM are adjusted. Next, a new simulation scenario is cr eated by loading the new lane-changing model to replace CORSIM original model. The new simula tion scenario is calibrated by using the same indices of measurement as in the previous calibration, and only the behavioral parameters within the LC plug-in are adjusted. By the end of this step, two ca librated simulation scenarios are obtained. One is with the original CORSIM lane -changing model, and the other is with the new lane-changing model. In the second step, multiple simulation runs for both calibrated models are conducted with OD demands measured from different day. Special MOEs (measurement of effectiveness) for lane-changing modeling, such as lane-based averag e speed, lane-based travel time, vehicle lane distribution and number of lane changes by vehi cle, are obtained to compare with the field observations. Results from the newl y developed model are expected to have a better match to the

PAGE 68

68 field data. Otherwise, the new model should be tuned up until the major ity of results are not inferior. Goodness-of-fit statistics may be used to evaluate the effectiv eness of the new model quantitatively. I. Model Implementation and Calibration II. Model Validation Figure 3-5. Procedures include d in the systematic validation

PAGE 69

69 Finally, sensitivity analyses are conducted for the simulations with new and original lane-changing models. Field traffic flows under different ti me periods are measured as the inputs. Various MOEs for lane-changing modeling are ob tained from the two models, and compared with each other, along with the field observations measured from same traffic conditions. It is anticipated that the results from the existing lane-changing model in CO RSIM would serve as a benchmark/test bed for the improvement of simulation capability offered by the new lanechanging algorithm. 3.5 Summary and Conclusions A general methodology and implementation framework for modeling lane-changing behaviors based on driver character istics and field data have been presented in this chapter. Two experiments, focus group study and i n-vehicle data collection, ar e designed to obtain the lanechanging related driver charac teristics and field maneuver da ta. With the inner connection attained from the experiments, a probabilistic lane-changing reason model and a gap acceptance model can be developed based on the empirical data. Strategies of implementing the new lanechanging model in CORSIM are also presented. Various calibration a nd validation endeavors can be included to test the simulation capabilities of new model. Structurally, the proposed met hodology provides a framework that can be used to modeling other driver behaviors (such as car-following) on urban arterials or the behaviors on other types of facilities besides/in addition to urban arterial s. It can be viewed as a hybrid of the human behavior research modeling extended to the transportation settings.

PAGE 70

70 In accordance with the requirements of the Univ ersity of Floridas Institutional Review Board (IRB), all research involvi ng human subjects needs to be approved by the relevant IRB Office prior to conducting any activities. Th e materials that are submitted to IRB-02 (UF Campus/Non Medical) for the experiments in th is research (focus group study and in-vehicle data collection) are provided in APPENDIX A

PAGE 71

71 CHAPTER 4 FOCUS GROUP-BASED STUDIES In Chapter 3 the research procedure was divi ded into four steps, and the motivation and anticipated results for each step were discusse d. This chapter presents the focus group study conducted in this research to obtain personal perceptions and at titudes regarding lane-changing maneuvers. The main objective is to use focus groups to obtain driver behavior-related data that can be used to model lane changes in an urban street environment. More specifically, the three sub-objectives of this research are: To develop an appropriate classification scheme for dr iver types based on driver background information (such as age, gender, etc.) and responses during the focus group discussion; To obtain the likelihood of attempting a given discretionary lane change (DLC) for different types of drivers; and To determine factors and parameters affecti ng the execution of a particular lane changing maneuver (mandatory or discretionary ) as a function of driver type. The remainder of the chapter is structured as follows. Section 4.1 presents the preparation and implementation of the focus group experi ment. Typical lane-cha nging scenarios are examined to obtain the level of likelihood in changing lanes, as well as important factors participants identified to affect their lane-c hanging behaviors. Next in Section 4.2, the quantitative and qualitative results for the fo cus group discussion are analyzed, followed by possible relationships between driv er behavior and driver charact eristics. Finally, the chapter ends with a summary and conclusion s of the study in Section 4.3.

PAGE 72

724.1 Focus Group Preparation and Implementation This section presents the focus group implemen tation details. First, the questions for the focus group discussion are provide d. Next, participants recrui tment and prescreening-related procedures are presented. Finally, moderation issues, including definitions and others are provided. 4.1.1 Preparation of Questions Developing good questions is very important for the focus groups study, which helps to generate desirable and useful results. Good questions includ e both a good questioning route and the effectiveness of each question. Krueger a nd Casey (2000) concluded the qualities of a good questioning route for focus groups st udies are: 1) having an easy beginning; 2) being sequenced; 3) moving the topic from general to specific; and 4) using the time available wisely. In the same reference, the qualities of good questions for focu s groups studies are desc ribed as: 1) including good directions; 2) one-dimensional; 3) open-ended; 4) short but clear; 5) easy to say; 6) adopting words participants would use; and 7) sounding conversational. Each quality is explained in detail in the corresponding chapte r of their book. This se ction introduces how the question route and questions for this research were prepared. Typically, a focus group discussion includes ab out 10-12 questions within two hours, and each question may function differently in the que stioning route to facilitate the moderating process. Krueger and Casey (2000) divided the categories of ques tions according to the purpose, in which an effective question route may in clude open questions, introductory questions, transitory questions, key questions and ending questions.

PAGE 73

73 Opening questions are ice-breaker questions, wh ich get every participant to talk early and help them to feel comfortable. The regular time is less than 30 seconds per person, and its important that the question does not highlight po wer and status differences among participants. In this study, the opening question (Table 4-1 Q1 ) is planned to let the participants introduce themselves, as follows: Tell us who you are, an d do you enjoy driving? Why? The overall time for this question is scheduled as about 3 minutes. The introductory question introduces the topic of discussion and help s people start thinking about their connection to the topi c of lane changing. Generally, they are open-ended questions which encourage conversation on the understanding of particip ants. In our research, one introductory question (Table 4-1 Q2) is: Wha t comes to your mind when you hear about the term change lanes? The overall time is scheduled as 5-6 minutes. The transitory question (Table 4-1 Q3) moves the conversation from general lane changing to the key content of the study, the potential reasons that may invoke a lane change, and made the participants aware of how ot hers view the topic. During th is question, participants are becoming aware of how others view the topi c, and refresh their thoughts from another perspective. Although the introductory question su rfaces the topic of di scussion, it is the transition question which makes rea l connection between the participants and the topic. In this research, such a question is proposed as: Think about the reasons invoking a lane change. Do you consider there are many differences between you and other drivers?

PAGE 74

74Table 4-1. Focus group categories and questions Q1 Opening Question Tell us a few things about yourself. Do you enjoy driving? Do you spend a lot of time on driving? How long have you been drivi ng? Q2 Introductory Question What comes to your mind when you hear lane change? Q3 Transition Question Are there any differences between the way you change lanes and that of other drivers (on urban streets)? Q4 Key Questions (Category 1: Likelihood of a Lane Change Discussion) Suppose you are driving on an urban street with three lanes. Pl ease evaluate how likely you are to conduct a lane change for ea ch given DLCs (the proposed list is given out in Table 2). Pl ease list any additional reasons you may have thought of. Q5 Key Questions (Category 2: Significant Factors in Lane Changing) In this category, a set of lane-changing scenario is given. Please identify factors affecting the manner and timing of your dec ision. Sce.1 Upcoming left/right turn: a. Left turn situation, b. Ri ght turn situation, c. Any di fferences between two turnings? Sce.2 Current lane is not available downstream: a. Road incident b. Work zone, c. Change in channelization of the current lan e Sce.3 Stopped bus at bus-stop: When driving on your lane, a bus (city bus, not school bus) in front is loading/unloading passe ngers Sce.4 Another vehicle merges into your lane: When driving in your lane, another vehicle is attemp ting to enter into your lane. Sce.5 Slow moving vehicle: When driving in your lane, the vehi cle in front of you is driving slower than you would like. Sce.6 Queue length advantage: When approaching an intersection, th e queue in your lane is longer than that of other lanes. Sce.7 Truck/heavy vehicle influence: There is a truck/HV in front of you blocking your view, and is traveling at desired speed. Sce.8 Tailgating by another vehicle: When driving in the center lane, you find that the vehicle behind you is tailgating you. Sce.9 Pavement condition: When driving in your lane, you find th e other lanes of the road have better pavement conditions. Q6 Key Questions (Category 3: Vehicle Interaction Discussion) In the next few slides, I will ask you about your actions dur ing a lane-changing maneuver assuming the traffic is congested. Pl ease describe your thoughts in planning and completing the maneuver. Sce.1 You need to change lanes: a. You are planning to merge to the curb-side lane; b. You are planning to merge to the middle lane. Sce.2 The other vehicle is changing lanes: a. The other vehicle is atte mpting to merge to the curb-side lane in front of you; b. The other vehicle is planning to merge to the median-side lane in front of you; c. Any differences between the two scenarios? Q7 Ending Question: Today, we began with the major possible re asons that would invoke a lane change a nd the level of likelihood in executing it. Ne xt, for each reason, the major effective factors which affect drivers decisi ons regarding lane changing were enumerated and examined. Final ly, the possible driver interactions involved in a la ne change behavior were discussed. Is ther e anything you want to say but didnt ge t a chance?

PAGE 75

75 This question first moves the conversation from th e general lane-changing to the topic of lanechanging reasons, and then it makes the participants become aware of how others view the topic. The scheduled time is 5-8 minutes. Key questions are the ones that require the grea test attention. Three key questions (Table 41Q4 to Q6) were used in this study. The first one (Table 4-1 Q4) star ts with introducing the definitions of MLC and DLC, and a list of DLC reasons is provided to participants. Next, they were asked to choose the level of likelihood that they may change lanes for each reason listed based on their driving experien ces. Participants were encour aged to add new lane-changing situations to the pre-selected list, and answers from each participant were recorded. The form used to assess DLC reasons is shown in Table 4-2. Five levels of likelihood for attempting a lane change for a given reason were defined. During th is part of the discussi on, only the discretionary lane changes (DLCs) were cons idered since the likelihood of a ttempting to change lanes for mandatory lane changes (MLCs) would be close to 100%. The output of this discussion is a comprehensive list of reasons, along with the level of lik elihood that a participant would change lanes for each of the identified reasons. The second key question (Table 4-1 Q5) de monstrates example scenarios for each particular lane change. Participants are asked to discuss and describe their behaviors under each lane-changing scenario (for both DLC and MLC), so that the major factors that affect their decision on attempting a lane change can be obtained.

PAGE 76

76 Table 4-2. Form for documenting th e level of likelihood for DLC reasons aLevels of Likelihood List of Discretionary Lane-Changing (DLC) Scenarios Lev. 1 Lev. 2 Lev. 3 Lev. 4 Lev. 5 R1. Change lanes to pass a stopped-bus at a bus stop R2. Change lanes to allow a vehicle to merge into your lane R3. Change lanes to pass a slower moving vehicle R4. Change lanes when the line of queuing vehicles is shorter in other lanes R5. Change lanes because there is a heavy vehicle/truck in front of you R6. Change lanes to avoid a vehicle tailgating you R7. Change lanes due to pavement conditions; R8. Other reason(s) please specify Notes:a Definitions of the levels of likelihood Lev. 1: Generally do not conduct (< 10%, weak) Lev. 2: Sometimes conduct but more likely do not (10% 40%) Lev. 3: Sometimes conduct, and sometimes do not (40% 60%) Lev. 4: More likely conduct (60% 90%) Lev. 5: Generally conduct (> 90%, strong) Figure 4-1 presents the sketches used to de scribe each scenario to the participants.

PAGE 77

77 (a) (b) Figure 4-1. Typical lane-changi ng scenarios occurred on urban stre ets (a) Lane change for the upc oming right/left turn (b) Upc oming lane is not available (c)Stopped bus at bus-stop (d) Another vehicle merges into your lane (e) Slow moving vehicle (f) Queue length advantage (g) Truck/heavy vehicle influence (h) Tailgating by a nother vehicle (i) Pavement conditions

PAGE 78

78 (c) (d) (e) (f) (g) (h) (i). Figure 4-1. Continued

PAGE 79

79 The output from this question is a list of f actors affecting the like lihood of changing lanes for a particular reason. An example is presented in Table 4-3 for the work z one lane closure, with factors and their respective importance from a pa rticular driver. The factors were identified by this participant as: lane-changing by the front vehicles, congesti on on the current/target lane and the presence of worker and m achinery. The corresponding levels of importance are: very important, important and not so important. The obj ective of this discussion is to identify the factors affecting lane-changing beha vior under different scenarios, a nd link the driver type to the factors considered in each lane-c hanging situation. The third key question (Table 4-1 Q6) aims to capture the interactions (cooperation and co mpetition) among drivers during lane-changing maneuvers. In this question, two directions of maneuvers (merging toward the curb-side lane or the median-side lane) were investigated. The participants were asked to describe their actions under each type of maneuver assuming that the traffic is congested. The objective of this part of discussion is to unveil possible driver interactions affecting lane changing, so that new algorithms can be developed to model these inte resting behaviors. Endi ng questions intent to close the discussion and enable participants to reflect on previous comments. The ending question (Table 4-1 Q7) of this research was designed as: Today, we began with the major possible reasons that would invoke a lane change and the level of frequency for executing it. Then for each reason, the major effective factors which affect drivers decision on lane change were enumerated and examined. Finally, the possi ble interactions involv ed in a lane change behavior were discussed. Did I co rrectly describe what was said here? Is there anything you want to say but didnt get a chance?

PAGE 80

80 The detailed lists of questions and figures used in the focus groups study are provided in Table 4-1 and Figure 4-1, with the modera ting scripts are presented in APPENDIX H. Table 4-3. Factors and the respective im portance for a given lane-changing situation Q5.2.b) Work zone lane closure; Very important Important Not so important Factor 1: lane-changi ng by the front vehicles Factor 2: congestion on current/target lane F actor 3: presence of worker and machinery Factor 4: 4.1.2 Participant Recruitme nt and Prescreening In accordance with Institutional Review Board (IRB) requirements, all research involving human subjects needs to be approved by the relevant IRB Office prior to conducting any activities. An application was submitted to IRB-02 (UF Campus /Non Medical) for this study in Dec. 2007, and a formal approval was obtained in Jan. 2008. The advertisement for recruitment was posted at public locations including the University of Florida campus, Gainesville downtown transit transfer stati on, Alachua county library and several supermarkets (APPENDIX E). In addi tion, the advertisement was placed on the classifieds in the local newspaper (Alligator), and sent to the UF ASCE student chapter and UF graduate students and staff. A web page was created and posted on the project website ( http://grove.ufl.edu/~jiansun). A prescreening procedure was de signed to obtain age, gender, race, resid ence, drive experience, and vehicle ownership information. Respondents could choose to complete the prescreening questionnaire and submit answers onlin e, or download the questionnaire from the server and re spond offline through email or mail.

PAGE 81

81 For lane-changing behavior, ther e is no special information-ri ch participant, and no such focus group study has really been conducted previously. Consequently, no specific criteria were used in the prescreening procedure. Any person with a valid driver license can be considered as a qualified candidate. Two other general criteria for participants recruitment were set as: 1. must have driving experience no less than three years; 2. should have liv ed in Gainesville, FL, for at least one year. The discussion time and the compensation were set as two hours, $50 pe r participant. By the end of the recruitment, responses from 84 pa rticipants were received. Previous focus groups studies in the transportation area (Loukopoulos et al., 2004; L oukopoulos, 2005) indicated that a candidate prescreening procedure is important for the quality of final results. The prescreening questionnaire (APPENDIX F) helped to identify qualified participants and collect useful background information. Participan t selection was based on age, ge nder, driving experience, and vehicle ownership to ensure a diverse group of participants. A total of 21 participants were invited to join the three focus groups. Four of these participants unexpect edly didnt attend. Two of the sessions had six participants, and the th ird had five. The detailed background information of the participants in these three gr oups is presented in Table 4-4.

PAGE 82

82Table 4-4. Personal background informa tion of the focus group participants ID Gender Age Group Experience Occupation Driving Frequency Hours Per Week Peak/ Nonpeak Vehicle Ownership 02-01 Female 20-29 3-9 years UF undergradua te Everyday < 4 hours Peak Sedan/coupe 02-03 Female 30-39 > 10 years Promotion representative Everyday 8-14 hours Peak Sedan/coupe 02-04 Male 50-59 > 10 years Film maker So metimes 4-8 hours Non-peak Sedan/coupe 02-05 Male 20-29 3-9 years Undergraduate Everyday 8-14 hours peak Sedan/coupe 02-06 Male 30-39 > 10 years UF graduate Usually 4-8 hours Peak Sedan/coupe 03-01 Male 30-39 > 10 years UF undergradua te Sometimes 4 hours Peak Pickup/SUV 03-02 Male 20-29 3-9 years Just graduated Everyday 8-14 hours Peak Sedan 03-03 Female 30-39 3-9 years UF graduate Everyday 4-8 hours Peak Sedan/coupe 03-05 Male 50-59 > 10 years Vocation instructor Everyday 4-8 hours Peak Sedan/coupe 03-06 Female 20-29 > 10 years Truck driv er Everyday > 14 hours Any TimeTruck 03-07 Male 40-49 > 10 years Shop owner Usually > 14 hours Peak Pickup/SUV 04-01 Female 30-39 > 10 years Secretary Everyday 8-14 hours Peak Sedan/coupe 04-02 Male 20-29 3-9 years UF undergraduate Sometimes 4-8 hours Peak Sedan/coupe 04-04 Male 20-29 > 10 years UF graduate Everyday 4-8 hours Peak Sedan/coupe 04-05 Female 20-29 3-9 years UF graduate Everyday 8-14 hours Any TimePickup/SUV 04-06 Male 30-39 > 10 years Pizza delivery Everyday > 14 hours Any time Sedan/coupe 04-07 Male 40-49 > 10 years Fitness traine r Sometimes 8-14 hours Peak Pickup/SUV

PAGE 83

83 4.1.3 Other Issues Upon arrival and before the discussion, a check-in procedure was followed and each participant was asked to 1) show their drivers license for identification, 2) sign the informed consent form (APPENDIX C) and 3) complete a background survey by answering six multiplechoice driver habits-related questions (A PPENDIX G). On the informed consent form, participants were fully briefed about the objec tives of the experiment. The discussion of each focus group was audio-taped with the permission of the participants (Washburn and Ko, 2007). Since it is difficult to anticipate the implementation of results of focus groups during the planning, a small pilot focus group study was pla nned and carried out in advance. The proposed pilot participants consist of faculty, staff and st udents from the University of Florida. Starting from the background survey and participants check-in, the whole procedure and moderating questions were the same as those enumerated for the real focus groups study. With the pilot study, defects within the curre nt questions and moderating scripts were exposed, so that significant flaws could be avoided during the real discus sion. However, results from this pilot study were not included in the further analysis. Three focus group discussions were conducted from April to July, 2008. By studying and comparing answers from each participant, it was found that the range of ideas werent getting new information, which is referred as the reach of saturation point (Morgan, 1997). As a result, no additional focus groups were needed. The reason for planning multiple groups is because focus groups are analyzed across groups, so that the patterns and themes can be obtained across groups.

PAGE 84

84 4.2 Analysis of the Results The information obtained directly from the fo cus groups are: 1) participants personal background information; 2) likelihood level of ch anging lanes for each driver and for each DLC reason; 3) factors that affect the lane-changing maneuver for th e situations examined; and 4) driver interactions that may occur when changi ng lanes. Various analyses were conducted to obtain information related to the research objectives. First, the drivers were classified into groups used cluster analysis (Tibshiran i et al., 2001). Second, for each of the groups identified using cluster analysis, the probability of various actions was obtained. Lastly, the critical factors affecting lane changing for each lane-changing scen ario were identified. The detailed description of each procedure is provided below. 4.2.1 Driver Type Classification Scheme To classify drivers into groups, first the driver background information (driver aggressiveness) was used to divide the participants into groups/clusters. Then the overall intracluster variance on the likelihood of changing la nes for each scenario is calculated and aggregated to select the most a ppropriate number of groups that s hould be used in this study. The number of groups was further confirmed qualitative ly based on verbal expressions obtained from the focus group discussions. Table 4-5 presents the driver aggressiveness reported by each participant along with their corresponding likelihood of executing a lane change, as spec ified by each participant during the focus group discussions (Table 4-1 Q4).

PAGE 85

85Table 4-5. Driver-based likelihood of executing a discretionary lane change Aggressiveness (1-10) Discussion Resu lts (Level of Likelihood, 1 5) ID SelfFriends Overall R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 Avg. 02-01 7 8 7.5 4 3 5 5 5 2 5 3 3 3.8 02-03 6 7 6.5 1 3 5 5 5 1 5 5 3 3.6 02-04 2 4 3 2 3 4 3 5 3 2 4 4 3.5 02-05 6 5 5.5 4 3 5 4 5 1 4 4 4 3.9 02-06 5 6 5.5 1 3 5 4 5 2 3 3 5 3.5 03-01 5 7 6 2 3 4 2 4 2 3 5 4 3.4 03-02 6 8 7 5 1 5 4 5 3 4 4 4 4.0 03-03 7 6 6.5 5 3 5 4 5 3 4 5 5 4.3 03-05 5 5 5 4 4 3 1 3 3 1 5 3 3.0 03-06 5 6 5.5 5 4 4 4 4 2 3 4 4 3.8 03-07 6 6 6 4 3 5 4 4 2 5 5 3 3.9 04-01 5 6 5.5 5 4 3 4 4 3 5 5 4.2 04-02 5 5 5 5 4 4 3 5 3 4 5 4.0 04-04 3 3-4 3.25 4 3 3 4 4 2 5 4 3.7 04-05 6 6 6 4 5 5 4 3 1 4 5 3.9 04-06 6 7 6.5 1 5 5 4 1 1 2 5 3.2 04-07 5 6 5.5 2 5 4 5 4 2 4 3 3.8 Avg. 5.3 6.0 5.65 3.4 3.5 4.4 3.8 4.2 2.1 3.7 4.2 4.7 3.9 3.8

PAGE 86

86 The left part of the table provides the self-e valuation and the percei ved friends evaluation obtained from the background survey. An overa ll aggressiveness for each participant is calculated by averaging the self-evaluation and the friends evaluation values. The results show that the self-evaluation value is generally sligh tly less than the friends evaluation (Selfavg.= 5.3, Friendsavg.= 6). The right part of the tabl e provides the likelihood of changing lanes under various scenarios, as reported by each particip ant. In addition to the pre-selected list of discretionary lane-changing (DLC) reasons show n in Table 2, four other DLC reasons were proposed by the participants: R8: In a corridor with many traffic lights, changing lanes to avoid backed-up turning movements; R9: Changing lanes to avoi d scooters/pedestrians; R10: Changing lanes to avoid an erratic driver The level of likelihood to change lanes was averaged across all situations (from R1 to R10) for each participant. However, the relationship be tween the average level of likelihood to change lanes and drivers aggressiveness is not straightforward. For some situations, such as changing lanes to give way to merging vehicles (R2), a defensive driver may have a high probability of changing lanes, and a more aggressive driver may speed up rather than change lanes. To classify drivers into groups, the K-mean s algorithm (Kanungo et al., 2002) was used to cluster n (n = 17) participants ba sed on overall aggressiveness into k (k = 1, 2, 3, 4, or 5) partitions, k < n. The algorithm (provided in APPENDIX I) attempts to find the centers of natural clusters in a given data set, and assum e s that the drivers aggressiveness form a vector space. Eq. (4-1) gives the objective of the algorithm to minimize total intra-cluster variance over all partitions: k iS x ijijx V1 2)( min (4-1) where:

PAGE 87

87 xj is the overall aggressiveness level for each participant j, obtained from Table 4-5, k is the number of clusters with in each lane-changing situation, Si represents each cluster (i = 1, 2, k), and i is the centroid point of cluster i. In Eq. (4-1), each element xj is grouped to the cluster i which has a minimal distance from xj to its centroid i compared to the other cluster centroids. By setting the cluster number as 1, 2, 3, 4, and 5 respectively, centroids for the clusters were obtained as follows: 1) for cluster number = 1, centro id for each cluster is 5.73; 2) for cluster number = 2, centroids for each cluster are 4.68 and 6.57; 3) for cluster number = 3, centroids for each cluster are 3.1, 5.62, and 7.02; 4) for cluster number = 4, centroids for each cluster are 3.1, 5.2, 6.12 and 7.02; 5) for cluster number = 5, centroids for each cluster are 3.1, 5.2, 6.12, 6.78 and 8. Next, the reason-based likelihood in formation is used to decide the most appropriate cluster number. The overall intra-cluster variance on the level of likelihood for each lane-changing situation (as reported in Table 4-5) was calcula ted, and accumulated across all reasons using Eq. (4-2): 10 11 2)( )(R R k iS l ijijl kW (4-2) where: li is the level of likelihood to change lanes for situation i for participant j, k is the number of clusters with in each lane-changing situation, Si represents each cluster (i = 1, 2, k), and i is the centroid point of cluster i. The W value for each classifi cation was calculated as: )1( W161.92, )2( W154.59, )3( W145.14, )4( W137.04 and )5( W135.68. Since the clustering method aims to put participants into clusters accord ing to closest similarity rules, and no a priori hypotheses were made, statistical significance tes ting is not appropriate. Conseque ntly, intra-cluster dissimilarity

PAGE 88

88 (low values when the partition is good) is used to determine th e appropriate number of clusters. The Hartigan index (Tibshirani et al., 2001), wh ich indicates the dissim ilarity that will be removed by splitting the k clusters into k+1 clusters, was used: )1( )1()( *)1()( kW kWkW knkH (4-3) where: n is number of objects to be clustered, n = 17, k is the number of cl usters used, and W(k) is the value calculat ed from Eq. (4-2). Using Eq. (4-3), the indices for the number of clus ters equaling 2, 3, 4 and 5 were calculated as H(1) = 0.71, H(2) = 0.91, H(3) = 0.77 and H(4) = 0.12. A large descent is found to occur from H(3) to H(4), which means by splitting the 3 clusters into 4, the dissimilarity is removed largely, while by splitting into 5, the dissimilarity is no t removed as much. Figure 4-2 provides the results of analysis for number of clusters ranging from 1 to 5. When the cluster number is larger than 4, the intra-cluster dissimilarity does not decrease much. Therefore, it is recommended that the appropriate number of cl usters used is 4. Figure 4-2. Results for the clusteri ng with different number of cluster Using the overall aggressiveness obtained together with the above K-mean cluster analysis, the participants can be categorized into four gro ups defined as L1 (<= 4.1), L2 (4.2 5.6), L3

PAGE 89

89 (5.7 6.5) and L4 (>= 6.6). In the remainder of this section, th is grouping is further linked to verbal expressions the participan ts used during the focus groups. During the focus group discussion, for the set of discretionary lane changes on urban streets, participants were asked to explain their thought process and how they would behave in each situation. The responses, along with the extent of the risk they would take and how much they focus on themselves vs. on surrounding vehicles, were evaluated. It was found that four elements were frequently considered by the participants: desi rability of speed advantage, tendency for risk taking, consideration of consequences and degr ee of selfishness. Based on verbal expressions on these four elements, the following types of drivers were identified: Type A: Drivers would not change lanes for most situations. They always want to keep the current lane because they are risk averse. Meanwhile, this type of drivers always considers other vehicles and would likely give way to vehicles merging into the current lane, or try not to block others. They can be described as the least aggressive drivers. Type B: Drivers would like to get a better positi on or speed advantage for some situations (the number is close to 4) under very low risk, but wouldnt on ot hers. They mentioned more details such as landscape or pictures and bumper stickers on the front vehicles in deciding whether to change lanes. This type of driver says lane changing depends on their mood: they would likely slow down if they ar e not in a hurry. Compared to the previous group, drivers in this group have more will ingness to change lanes to obtain speed advantage, and consequently are somewhat mo re aggressive. However they generally dont like to take risks. Type C: Drivers aim to get a better position or sp eed advantage if they have a chance. However, they would also consider other factor s, such as traffic congestion or destination, which they consider more important than speed advantage and better positioning. Compared to the previous driver group, this group of drivers is more ambitious in getting speed advantage, and would take in creasing risks in changing lanes. Type D: Drivers in this group would always try to get a better position or speed advantage whenever they have a chance. They barely think about other drivers. Position and speed are their first consideration. They would change lanes without an y hesitation, and would risk without caring much about the en vironment or other drivers. The categorization, with the corresponding observati on of characteristics of the participants, is summarized in Table 4-6. It was found that some of the four driver characteristics factors

PAGE 90

90 correlate highly to each other. For example, th e high risk-taking drivers generally desire speed advantage highly. Table 4-6. Driver type categorization by the ch aracteristics demonstrated in verbal expression Driver Type Desiring Speed Adv. Risk Taking Consideration of Consequence Selfishness Type A No No No No Type B Sometime No Yes, always No Type C Yes, always Sometime Sometime Yes, always Type D Yes, always Yes, alwaysNo Yes, always By tagging each participant w ith a corresponding driver type (A, B, C or D), it was found that for most of them, the driver type inform ation obtained from the focus group discussion is consistent to the groups defined in the K-mean cluster analysis (L1 to Type A, L2 to Type B, L3 to Type C, and L4 to Type D), as shown in Table 4-7. Table 4-7. Consistency between the cluste ring result and driver type demonstrated ID Overall Aggressiveness K-Mean Cluster Type Agree or not 02-01 7.5 L4 D Y 02-03 6.5 L3 C Y 02-04 3 L1 A Y 02-05 5.5 L2 B Y 02-06 5.5 L2 C N 03-01 6 L3 C Y 03-02 7 L4 D Y 03-03 6.5 L3 C Y 03-05 5 L2 B Y 03-06 5.5 L2 B Y 03-07 6 L3 C Y 04-01 5.5 L2 C N 04-02 5 L2 B Y 04-04 3.25 L1 A Y 04-05 6 L3 C Y 04-06 6.5 L3 C Y 04-07 5.5 L2 B Y

PAGE 91

91 The only exceptions are participan ts 02-06 and 04-01. Both of these participants were in the aggressiveness group L2, with an overall aggressiveness value as 5.5. However, they were assigned to type C instead of type B since they demonstrated amb ition in getting speed advantage during the discussion, and generally wont give way to more than one vehicle. This is probably due to the fact that gr oups L2 and L3 are rather close to each other. However, since only 2 of 17 samples do not agree, it was deemed reasonable to divide the drivers into four groups as recommended above. 4.2.2 Probability of Various Action s for Different Driver Types The classification of drivers developed in the previous section was next applied to each of the lane-changing situations. The num ber of drivers in each group (L1, L2, L3 and L4) is 2, 6, 7 and 2 respectively. Table 4-8 presents the fr equency of likelihood level for each lane-changing maneuver by driver type along with the mean and the standard deviation. For example, for R1 (change lanes to pass a stopped-bus at a bus stop ), the average acceptance levels for the four groups (L1, L2, L3 and L4) are 3, 3.6, 3 and 4. 5 respectively. This mean s that the drivers in group L4, which are more aggressive, are more like ly to change lanes to pass a stopped-bus. This is consistent with the expectation for such a maneuver. Similar trends can be observed for reasons R3, R4, R5, R7 and R8, where the more aggr essive drivers are more likely to make lane changes. For reason R2 (change lanes to allow a vehicle to merge into yo ur lane), the likelihood level for the four groups L1, L2, L3 and L4 are 3, 4, 3.5 and 2 respectively. In this case, the most aggressive drivers are typically not willing to allow the merge. Instead of giving way by changing lanes, they accelerate or at least maintain their speed to prevent the merging vehicle from changing lanes. Meanwhile, defensive drivers (L1) typically choose to decelerate instead of changing lanes. The other drivers (L2 and L3) are found with the larges t likelihood to change lanes in this situation. Reasons R6 and R10 have similar trends as R2. The other reason, R9

PAGE 92

92 Table 4-8. Lane-changi ng likelihood level for e ach driver group (L1-L4) Frequency by likelihood level Reasons Freq. Total # 1 2 3 4 5 Likelihood ( / ) L1 2 0 1 0 1 0 3.0 /1.4 L2 6 1 1 0 1 3 3.7 /1.8 L3 7 2 1 0 3 1 3.0 /1.6 R1 (Stopped bus) L4 2 0 0 0 1 1 4.5 /0.7 L1 2 0 0 2 0 0 3.0 /0.0 L2 6 0 0 1 4 1 4.0 /0.6 L3 7 0 0 5 0 2 3.6 /1.0 R2 (Vehicle merge) L4 2 1 0 1 0 0 2.0 /1.4 L1 2 0 0 1 1 0 3.5 /0.7 L2 6 0 0 2 3 1 3.8 /0.8 L3 7 0 0 0 1 6 4.9 /0.4 R3 (Slow vehicle) L4 2 0 0 0 0 2 5.0 /0.0 L1 2 0 0 1 1 0 3.5 /0.7 L2 6 1 0 1 3 1 3.5 /1.4 L3 7 0 1 0 5 1 3.9 /0.9 R4 (Queue advantage) L4 2 0 0 0 1 1 4.5 /0.7 L1 2 0 0 1 1 0 3.5 /0.7 L2 6 0 0 1 3 2 4.2 /0.8 L3 7 1 0 1 2 3 3.9 /1.5 R5 (Heavy vehicle) L4 2 0 0 0 0 2 5.0 /0.0 L1 2 0 1 1 0 0 2.5 /0.7 L2 6 0 3 3 0 0 2.5 /0.6 L3 7 4 2 1 0 0 1.6 /0.8 R6 (Tailgating) L4 2 0 1 1 0 0 2.5 /0.7 L1 2 0 1 0 0 1 3.5 /2.1 L2 6 1 0 2 2 1 3.3 /1.4 L3 7 0 1 1 3 2 3.9 /1.1 R7 (Pavement) L4 2 0 0 0 1 1 4.5 /0.7 L1 2 0 0 0 2 0 4.0 /0.0 L2 4 0 0 2 0 2 4.0 /1.2 L3 4 0 0 0 1 3 4.8 /0.5 R8 (Backup turning) L4 1 0 0 1 0 0 3.0 /0.0 L1 0 0 0 0 0 0 L2 2 0 0 0 1 1 4.5 /0.7 L3 3 0 0 0 0 3 5.0 /0.0 R9 (Pedestrian/scooter) L4 1 0 0 0 1 0 4.0 /0.0 L1 1 0 0 0 1 0 4.0 /0.0 L2 3 0 0 1 1 1 4.0 /1.0 L3 5 0 0 2 1 2 4.0 /1.0 R10 (Erratic driver) L4 2 0 0 1 1 0 3.5 /0.7 (changing lanes to avoid scooters/ pedestrians), does not have la rge differences across the four groups. Most drivers tend to change lanes for this particular r eason, no matter the driver type,

PAGE 93

93 which means the classification may not have significant impact on this situation. The comparative small sample size (6, instead of 17) for R9 is another possible cause for this similarity. In summary, likelihoods for lane changes are different depending on the scenarios. For some lane-changing reasons, aggressive drivers al ways have a higher like lihood to change lanes, while for other reasons they dont. Consequently, in modeling lane changes, both the driver type and the individual lane-changing scen ario should be considered. 4.2.3 Critical Factors for Each Lane-Changing Scenario Another important objective of the focus gr oup study is to obtain the significant factors considered by drivers for various lane-changing scenarios. These include all the DLC scenarios originally listed and the two MLCs (Table 4-1 Q5 Sce.1 9). The factors identified, as well as the driver-assigned importance leve ls (very important, importa nt, and not so important) were collected from each participant during the focus group discussions a nd aggregated by lanechanging scenario. A quantitative evaluation was ne xt designed by assigning each of three levels of importance (for very important important and not so impor tant as in Table 4-3) 9, 6 and 3 credits respectively. Only factors listed by the participants received credits. These credits were used only to provide a qua ntitative analysis base; any set of weights or credits could be assigned without alter the re sults f this analysis. Let Xij represent the credit(s) that participant j assigned to factor i then j ij iXS is the total credits earned by factor i from all participants. Thus Xij is as follows: by chosen not is 0 ;important" sonot as weighed and by chosen is 3 ;important"" as weighed and by chosen is 6 ;important" very as weighed and by chosen is 9jtparticipan ifactor jtparticipan ifactor jtparticipan ifactor jtparticipan ifactor Xij (4-4)

PAGE 94

94 By comparing the total credit earned by each factor i (i j ij iXS), the ones with higher credits were chosen by more participants, thus are more important. Table 4-9 presents the factors found to affect drivers lane-changing maneuvers for the lane-changing scenarios examined (see Table 4-1 Q5). Table 9 presents th e factors in decreasing or der of significance, as indicated by the participants. A discussion for each lane-changing scenario and the respective factors found to be importa nt is provided below: Scenario 1 Upcoming left/right turn (1a and 1b): The left turn and right turn scenarios were believed to be similar, and the factors found to be important for both are traffic congestion on the current lane and the target lane, the posted sp eed limit and travel speed, and the distance to the downstream intersection. The difference is that for the left turn scenario, drivers pay more attention to the traffic signal, while in the right turn scenario, the familiarity with road/alternative and the presence of pedestrian/bike la nes tend to affect drivers more. Scenario 2 Current lane is not available downstream (2a, 2b and 2c): The factors found to be important for these three scenarios are traffi c congestion on the current lane and the target lane, the posted speed limit and travel speed, and the distance to the lane termination location. The difference is that for road incidents, driv ers have to pay attenti on to oncoming emergency vehicles. In work zones, driver s pay attention to the signs an d the presence of workers or machinery. Scenario 3 Stopped bus at bus-stop: Four factors found to be impor tant are the distance to the bus, the traffic congestion/queue ahead, the numbe r of passengers at the bus stop and the location of the next bus stop. Other factors, such as pa ssenger loading stage and s ubject vehicle type were also mentioned. However, these factors were either not chosen by ma ny participants or not assigned as high a level of importance.

PAGE 95

95 Table 4-9. Important factors for individual lane-changing scenario Frequency Sce. Factors 3:weak 6:middle 9:strong Si: Total Credits Congestion on the left lane/easy to get a gap 2 5 7 99 Traffic signal/left turn signal 3 5 5 84 Speed limit/travel speeds 3 4 4 69 Distance to the intersection 4 5 2 60 Number of lanes 3 1 2 33 Vehicle type 1 1 1 18 Driver mood/in a rush 1 0 1 12 Pedestrian/scooter 0 0 1 9 1a. Left-turn Weather condition 0 1 0 6 Distance to the intersection 2 3 5 69 Speed limit/ travel speeds 4 3 4 66 Familiarity with road/alternative 3 5 2 57 Congestion on current and adjacent lane 1 2 4 51 Pedestrian/bike lane 5 3 1 42 Traffic signal/RTOR 3 1 2 33 Lane changes by other vehicles 2 1 1 21 Vehicle type 1 1 0 9 Sce. 1. Upcoming left/right turn 1b. Right-turn Weather condition 0 1 0 6 Congestion on current and merging lane 3 3 5 72 Lane changes by front vehicles 1 5 4 69 Distance to incident 3 2 4 63 Oncoming emergency vehicle 2 4 3 57 Relative speed 2 4 3 57 People in the roadway 1 2 1 24 Visibility 1 2 0 15 Debris in the roadway 1 1 0 9 Weather condition 1 1 0 9 2a. Road incident Vehicle type 0 1 0 6 Lane changes by front vehicles 3 5 4 75 Presence of worker or machinery 4 5 2 60 Congestion on current and merging 2 4 3 57 Number of lanes 3 2 4 57 Location of the first warning sign 3 2 3 48 Speed limit/travel speeds 2 0 1 15 Vehicle type 1 1 0 9 Weather condition 0 1 0 6 2b. Workzone Visibility 1 0 0 3 Level of congestion 2 5 7 99 Distance to the lane drop 3 6 5 90 Speed limit/travel speeds 4 7 2 72 Lane changes by front vehicles 1 2 1 24 Pedestrians/bikers 2 2 0 18 Vehicle type 1 2 0 15 Weather condition 0 1 0 6 Sce. 2. Current lane is not available downstream 2c. Lane channel. change Signal ahead 1 0 0 3

PAGE 96

96 Table 4-9. Continued Frequency Sce. Factors 3:weak 6:middle 9:strong Si: Total Credits Traffic congestion and queue ahead 2 3 6 78 Location of the next stop 1 5 3 60 Distance to the bus 1 3 4 57 Number of persons at the bus-stop 2 4 1 39 Stage of loading/unloading 3 2 0 21 Distance to my next turn 3 2 0 21 Speed limit/travel speeds 1 2 0 15 Mood/urgency 1 0 1 12 Vehicle type 1 1 0 9 Weekday or weekend 0 1 0 6 3. Stopped-bus Weather condition 0 1 0 6 Congestion on the target lane 2 7 4 84 Speed limit/travel speeds 2 5 4 72 Aggressiveness of the merge 3 6 2 63 Distance to my next turn 4 5 2 60 Merge and my vehicle type 1 4 3 54 With a turning signal or not 0 1 1 15 Direction of the merge 1 1 0 9 Presence of signal control 2 0 0 6 4. Vehicle merge Weather and my mood 1 0 0 3 Distance to my next turn 2 8 3 81 Speed limit/travel speeds 1 7 4 81 Congestion on target lane 1 5 4 69 My mood, hurry or not 5 4 2 57 Merge vehicle type 3 1 0 15 Presence of signal control/status 0 2 0 12 5. Slow vehicle Weather or visibility 0 2 0 12 Queue length difference 2 4 7 93 Distance to my next turn 4 6 3 75 Congestion on the target lane 1 4 3 54 Current signal status/green time 3 5 1 48 Vehicle type of the queuing vehicles 1 0 2 21 Number of lanes 2 1 1 21 Weather condition/visibility 0 2 0 12 6. Queue advantage My vehicle type 1 0 0 3 Travel speed/desired speed 1 5 6 87 Congestion on all lanes 1 5 4 69 Personally uncomfortable with HV 3 4 2 51 My vehicle type 3 3 2 45 Weather and visibility 2 3 1 33 Distance b/t mine and the HV 2 1 0 12 7. Heavy vehicle Time driving on current lane 1 1 0 9 Speed limit/travel speeds 2 5 6 90 Congestion on all lanes 3 4 5 78 The lane I am driving on 2 7 2 66 Vehicle type (mine and others) 5 2 2 45 Mood 3 1 1 24 Time driving on current lane 1 2 0 15 8. Tailgating Distance to my next turn 1 1 0 9 Distance to my next turn 3 3 5 72 How large the difference of the pavement 4 5 2 60 How long of the pavement diff. segment 6 3 2 54 Speed difference 4 1 3 45 Congestion on the other lanes 0 2 3 39 Weather and visibility 0 1 0 6 9. Pavement My vehicle type 1 0 0 3

PAGE 97

97 Scenario 4 Another vehicle merges into your lane: Five factors were found to be important. In addition to the posted speed limit, travel speed and traffic congestion on the target lane, the distance to the next turn, aggres siveness of the merge and vehicle types of the merger and the subject were also considered as important. Scenario 5 Slow moving vehicle: Most of the factors found to be important, such as traffic congestion on the target lane, the posted speed lim it and travel speed, and the distance to the next turn, are similar to previous lane-changing scenarios. Only one factor, the mood and urgency of the subject driver, is somewhat unique to this scenario. Scenario 6 Queue length advantage: Four factors were found to be important. In addition to the previously mentioned distance to the next turn and the traffi c congestion on the target lane, two other factors offered are the queue length di fference and the current signal status, both of which are more related to queuing. It was genera lly found that in most lane-changing scenarios, if there is a traffic signal in the vicinity, the si gnal status is always an important factor. Scenario 7 Truck/he avy vehicle influence: Four factors were found to be important: traffic congestion on all available lanes, the pos ted speed limit and travel speed, personally uncomfortable with HV, and the subject vehicle t ype. The latter two factors are related closely to the drivers line of sight. The participants e xpress more willingness to change lanes, when the line of sight is blocked. Scenario 8 Tailgating by another vehicle: Four factors were found to be important. First, the posted speed limit and travel speed, and traffic congestion on all av ailable lanes were considered important. Next, the participants also consider ed subject and follower vehicle types, and the subject lane (median or curb-side) as important. If the tailgating occurs on the median lane, they are even more willing to change lanes.

PAGE 98

98 Scenario 9 Pavement condition: Five factors were found to be important. As for the previous scenarios, the distance to the ne xt turn, speed difference and conge stion on the other lanes were considered important. Other factor s considered as important are the difference of the pavement quality and the length of the different pavement segment. As shown, some factors, such as congestion le vels and speed differences are common to all scenarios. Other factors applic able to particular lane change(s), were also found in each individual scenario. In addition to the factors identified by focus group participants as important, there are other types of data, such as vehicle acceleration/deceleration an d road geometry data, which are essential in lane-chang ing modeling. These need to be considered in conjunction with the driver-related factors identified in this analysis. Question Q6 (Table 4-1) focused on driver interactions that may occur during lane changing. The related conclu sions are provided below: In general, the variability of driver behavior when they are in the lag vehicle (with respect to the lane changing vehicle) is much larger than the variability of behavior in drivers making the lane change. Drivers in the lag vehi cle generally consider the traffic conditions when deciding how to react to a merging vehicle.. The type of driver in the lag vehicle is important. Under heavy traffic, lag drivers have to interact with merging drivers.. However, dr iver types A and B typi cally choose to give way and cooperate, while driver types C and D would accelerate or at least maintain their speed. The presence of intersections or drive ways is important to the lag vehicle drivers. They are more willing to cooperate if there are intersectio ns or driveways in the vicinity of the lane changing maneuver, since they consider that th e lane changing vehicle might need to turn soon. 4.3 Summary and Conclusions The lane-changing decision-making process on urban arterials depends largely on driver characteristics, which cannot be obtained from traditionally collected field data. In this research,

PAGE 99

99 a focus group study was conducted to document the drivers thinking process and perceptions related to lane-changing maneuvers. The followi ng conclusions were drawn from this study: Considering both personal background data a nd stated behavior data related to urban arterial lane-changing situations, the participatin g drivers were classifi ed into four groups using cluster analysis. Quantitative results from the questionnaire were found to be consistent to the qualitative verbal expression-based conclusions. The probabilities of changing la nes were obtained for each of the four driver type groups and for each lane-changing scenario. Factors affecting each lane-changing scen ario were obtained from the focus group discussion. Some of the factor s, such as congestion level and speed difference were found be apply to all scenarios, while several f actors were found to be unique to each lanechanging scenario. The results of this study can be implemented into micro-simulators to better replicate driver behavior in urban street networ ks: the classification into four driver types as well as their corresponding behavior can be used to model lane changes more accurately. For example, CORSIM currently allows the use of ten different driver types, however it is possible that only four driver types are needed to describe the tr affic stream accurately. These four driver types could be modeled to attempt and execute lane changes as described above, by considering the factors identified to be important to that group for each maneuver type.

PAGE 100

100 CHAPTER 5 IN-VEHICLE EXPERIMENT AND ANALYSIS In Chapter 4, the procedure and result analys is of the focus group study on lane-changing behaviors w ere presented. One of the issues with using the focus group study is that the research was based on focus group discussions and background surveys, wherein the participants may over-think their actions, while many actual drivi ng decisions were made instantaneously or in a very short time. Consequently a field experiment was designe d to observe and validate the stated driver preferences. This chapter describes the in-vehicle data collection experiment and the corresponding analysis procedure. The objective of this experiment is to observe the drivers action under various lane-changi ng scenarios, and to obtain th e quantitative values for the important factors identified dur ing the focus group study. The expe rimental results are used to develop new lane-changing models on urban arterials in further Chapter 6. This chapter is organized as follows: S ection 5.1 discusses the preparation and implementation of the in-vehicle experiment. Th e composition of the participating drivers, the testing routes and the techniques used during driving for this experiment are presented. Results from the in-vehicle data collection experiment are analyzed in Section 5.2. The quantitative values for the important factors in each lane-c hanging related maneuver were obtained from the in-vehicle video clips, so that further analys is regarding driver type classification can be conducted. The chapter ends with a summary of the findings from the in-vehicle experiment. 5.1 In-Vehicle Experiment Prep aration and Implementation During the in-vehicle experiment, an instrumented vehicl e was used to verify the lanechanging process and the gap acceptance charact eristics of a diverse group of drivers. Implementation details for the experiment are provi ded in this section. First, the participant

PAGE 101

101 characteristics are presented. Next, the route used in testing is provided. Finally, the driving test procedure, as well as recording techniques used during the in-vehicle expe riment, is discussed. 5.1.1 Participant Characteristics The same recruiting website and prescr eening procedure as the focus group study (http://grove.ufl.edu/~jiansun) we re used for participant recruitment in the in-vehicle experiment. From more than 150 responses, 40 driver s were invited with an intention to select a diverse group based on age, gende r, driving experience, occupati on and vehicle ownership. In addition, diverse drivers were purposely invited based on the indi cated aggressiveness. With an eye on verifying the results from the focus group study, ten of the subjects had participated in the focus group discussion. The compen sation was set at $50 per partic ipant. Table 5-1 presents a summary of the statistics related to age and gender for th e participants. The detailed background information about the participan ts is provided in Table 5-2. Table 5-1. Overview of the participants characteristics for the in-vehicle experiment Age and Gender 20 29 30 39 40 49 50 59 Total Male 6 (15) 5 (12.5) 6 (15) 6 (15) 23 (57.5) Female 4 (10) 6 (15) 3 (7.5) 4 (10) 17 (42.5) # of participant (percentage, %) Total 10 (25) 11 (27.5) 9 (22.5) 10 (25) 40 (100)

PAGE 102

102Table 5-2. Personal background information of the In-Vehicle experiment participants ID Gender Age Group Experience Occupation Driving Frequency Hours per Week Peak/ Non-peak Vehicle Ownership 05-01 Male 20-29 3 9 years UF undergraduate Usually 8 14 hr Peak Pickup/SUV 05-02 Male 40-49 > 10 years Shopper owner Usually > 14 hr Peak Pickup/SUV 05-03 Female 20-29 3 9 years Insurance company Everyday < 4 hr Peak Sedan/coupe 05-04 Male 50-59 > 10 years Film maker Some times 4 8 hr Non-peak Sedan/coupe 05-05 Female 30-39 > 10 years UF graduate So metimes 4 8 hr Non-peak Sedan/coupe 05-06 Male 20-29 3 9 years Medical studen t Usually 4 8 hr Peak Sedan/coupe 05-07 Female 30-39 > 10 years UF graduate So metimes 4 8 hr Non-peak Sedan/coupe 05-08 Female 30-39 > 10 years Advertisement promoter Everyday 8 14 hr Peak Sedan/coupe 05-09 Male 20-29 3 9 years UF undergraduate Sometimes 4 8 hr Peak Sedan/coupe 05-10 Male 30-39 > 10 years UF graduate Usually 4 8 hr Peak Sedan/coupe 05-11 Female 50-59 > 10 years Clerk at supermarket Sometimes > 14 hr Peak Sedan/coupe 05-12 Male 40-49 > 10 years Environment consultant Sometimes 4 8 hr Non-peak Sedan/coupe 05-13 Male 20-29 3 9 years Lawyer Ev eryday 4 8 hr Peak Sedan/coupe 05-14 Female 20-29 3 9 years UF undergraduate Sometimes 4 8 hr Peak Sedan/coupe 05-15 Female 50-59 > 10 years Unemployed So metimes < 4 hr Non-peak Pickup/SUV 05-16 Male 20-29 3 9 years UF staff Everyday 8 14 hr Peak Pickup/SUV 05-17 Female 20-29 3 9 years UF undergraduate Everyday 4 8 hr Peak Sedan/coupe 05-18 Male 30-39 > 10 years Body trainer So metimes < 4 hr Non-peak Sedan/coupe 05-19 Male 50-59 > 10 years Fire truck driv er Usually 4 8 hr Non-peak Mini-Van 05-20 Male 40-49 > 10 years Post-doctor researcher Usually < 4 hr Non-peak Truck

PAGE 103

103Table 5-2. Continued ID Gender Age Group Experience Occupation Driving Frequency Hours per Week Peak/ Non-peak Vehicle Ownership 05-21 Male 30-39 > 10 years Translator Sometimes < 4 hr Non-peak Sedan/coupe 05-22 Female 50-59 > 10 years Unemployed Everyday 8 14 hr Non-peak Jeep 05-23 Male 20-29 3 9 years UF graduate Everyday 8 14 hr Peak Sedan/coupe 05-24 Male 30-39 > 10 years Biology consultant Everyday 4 8 hr Peak Sedan/coupe 05-25 Female 30-39 > 10 years Secretary Everyday 8 14 hr Peak Sedan/coupe 05-26 Male 50-59 > 10 years Vocation consultant Everyday 4 8 hr peak Sedan/coupe 05-27 Male 40-49 > 10 years Veteran, resell desktops Usually 4 8 hr Non-peak Pickup/SUV 05-28 Male 40-49 > 10 years Small business owner Sometimes 4 8 hr Non-peak Pickup/SUV 05-29 Male 30-39 > 10 years Chef and painter Usually < 4 hr Non-peak Jeep 05-30 Female 30-39 > 10 years Criminal lawy er Everyday 4 8 hr Peak Sedan/coupe 05-31 Male 50-59 > 10 years Construction worker Everyday 8 14 hr Non-peak Jeep 05-32 Male 50-59 > 10 years Retired, consultant Sometimes < 4 hr Non-peak Pickup/SUV 05-33 Female 50-59 > 10 years School instructor Sometimes 4 8 hr Peak Sedan/coupe 05-34 Female 30-39 > 10 years Urban planne r Everyday 4 8 hr Peak Sedan/coupe 05-35 Female 40-49 > 10 years UF staff So metimes < 4 hr Non-peak Sedan/coupe 05-36 Male 40-49 > 10 years Real estate seller Everyday 4 8 hr Non-peak Truck, sedan 05-37 Male 50-59 > 10 years College instructor Sometime < 4 hr Non-peak Sedan, scooter 05-38 Female 40-49 > 10 years Part-time job Everyday 4 8 hr Peak Mini-Van 05-39 Female 20-29 3 9 years UF undergraduate Everyday 48 hr Non-peak Sedan/coupe 05-40 Female 40-49 > 10 years Part-time job Everyday 8 14 hr Non-peak Truck

PAGE 104

1045.1.2 Testing Route In addition to the participants recruitment, another important issue during the preparation is the selection of the driving route. The road segments selected were located in the city of Gainesville, FL, and each may invoke one or more la ne-changing scenario(s), such as left turn, right turn, work zone, stopped-bus, right-lane merging, and so on. By connecting these segments, two routes, shown in Figures 5-1 and 5-2, were es tablished for different ti me-of-day traffic data collection (with different levels of congestion).

PAGE 105

105 Figure 5-1. Proposed route for the field data collection (Newberry Road route)

PAGE 106

106 Figure 5-2. Proposed route for the fiel d data collection (Waldo Road route)

PAGE 107

107 Figure 5-1 presents the Newberry Rd route, which was followed in the early afternoon (3:15 pm to 4:20 pm). The total distance is about 16 miles, and the estimated non-congested travel time is about 40 50 minutes. Participants were asked to stop at check points during the test to discuss the lane-changing related maneuvers that occurred on the road, so that driver behavior related information can be better unde rstood. Three check points were selected: the OConnell center parking garage, Butler plaza and Oaks mall. By including the discussion time in each check point, which was set as 3-5 minutes each, the total experiment time was close to or slightly more than one hour. The various antici pated and potential lane-changing situations, with detailed information for check points and segments of the route, are listed in Table 5-3. The index number of each route segment is posted on Figure 5-1. Figure 5-2 presents the Waldo Rd route, which was used dur ing the PM peak hour (4:30 pm to 5:30 pm). The total distance is about 14.5 miles, and the estimated travel time for PM congested traffic is about 50 minutes. Three chec k points were selected: the Coastal Eng. lab, Butler plaza and NE 16 Ave. gas station. By in cluding the discussion time in each check point, which is set as 3-5 minutes each, the total experiment time is s lightly more than one hour. The various anticipated and potential lane-changing situations, with detailed information for check points and segments of the route, are listed in Table 5-4. The index number of each route segment is posted on Figure 5-2.

PAGE 108

108Table 5-3. Detailed route information for the In-Vehi cle data collection experime nt (Newberry Road route) Segment Index No. From To Turning at end of the segment Distance (mile) # of lanes Anticipated scenarios* Potential scenarios* Check point 1 Start Point: OConnell Center Parking Ga rage (University Ave. & Lemerand Dr) Segment 1 OConnell Museum Rd. (SW 8th Street) Left turn 0.3 1 N/A N/A Segment 2 Museum Rd. (SW 8th Street) Newell Dr. Right turn 0.4 1 N/A N/A Segment 3 Newell Dr. SW Archer Rd. Right turn 0.4 2 R1, R2, R3, R4 R8 Segment 4 SW Archer Rd. Butler Plaza (SW 37th Blvd) Right turn 3.2 2 R1, R2, R3, R4, R6 R8 Check point 2 Butler Plaza (SW Archer Rd. & SW 37th Blvd) Segment 5 Butler Plaza (SW 37th Blvd) SW. 34th Street Left turn 0.8 3 R1, R2, R3, R4, R6, R9 R8 Segment 6 SW. 34th Street Newberry Rd. (W. Univ. Ave.) Left turn 1.8 3 R2, R3, R4, R6, R7 R8 Segment 6 Newberry Rd. (W. Univ. Ave.) NW 8th Ave Right merge 1.9 2 R4, R6, R8 Segment 8 NW 8th Ave Oaks Mall Left turn 1.1 2 R1, R2, R3, R4, R7, R9 R8 Check point 3 Oaks Mall (6419 W Newberry Rd.) Segment 9 Oaks Mall NW 8th Ave Left turn 1.1 2 R2, R3, R4, R6, R7, R8 Segment 10 NW 8th Ave W. Univ. Ave. 1.9 2 R2, R6, R8 Segment 11 W. Univ. Ave. OConnell Right turn 1.6 2 -> 1 R2, R3, R4, R6, R10 R8 Check point 4 End Point: OConnell Center Parking Garage (University Ave. & Lemerand Dr) *Note: Abbreviations for Anticipated & Potential Scenarios R1Upcoming left/right turn at the immediate/next downstream intersection; R2Current lane is not available downstream (e.g. road incident, work zone or change in channelization of the current lane); a nd R3Passing a stopped-bus at bus stop; R4Giving way to a merging vehicle; R5Gaining speed advantage by overtaking a slower moving vehicle; R6Gaining queue advantage; R7Avoiding a truck/heavy vehicle influence; R8Avoiding the pressure imposed by tailgating; R9Attracted by a better pavement condition;

PAGE 109

109Table 5-4. Detailed route information for the In-Ve hicle data collection experi ment (Waldo Road route) Segment Index No. From To Turning at end of the section Distance (mile) # of lanes Anticipated scenarios* Potential scenarios* Check point 1 Start Point: Coastal Eng. Lab (1300 SW 6th Street) Segment 1 SW 6th Street SW16th Ave. Right turn 0.2 2 N/A N/A Segment 2 SW 16th Ave. SW 13th Street Right turn 0.6 2 R1, R2, R3 R8 Segment 3 SW 13th Street SW 8th Ave. (Museum Rd.) Left turn 0.5 2 R1, R2, R3, R4 R8 Segment 4 SW 8th Ave. (Museum Rd.) Newell Dr. Left turn 0.2 2 N/A R8 Segment 5 Newell Dr. SW Archer Rd. Right turn 0.4 2 N/A N/A Segment 6 SW Archer Rd. Butler Plaza (SW 37th Blvd) Right turn 3.2 3 R1, R2, R3, R4, R6 R8 Check point 2 Butler Plaza (SW Archer Rd. & SW 37th Blvd) Segment 7 Butler Plaza (SW 37th Blvd) S. Main Street Left turn 3.8 3 -> 2 R1, R2, R3, R4, R6, R9 R8 Segment 8 S. Main Street E. Univ. Ave Right turn 1.2 2 R2, R3, R4, R6, R7 R8 Segment 9 E. Univ. Ave. NE Waldo Rd. Left turn 0.8 2 R1, R2, R3, R4, R7, R9 R8 Segment 10 NE Waldo Rd. NE 16th Ave. Left turn 1.2 2 R2, R3, R4, R5, R6, R7, R9, R10 R8 Check point 3 Gas station (NE Waldo Rd. & NE 16th Ave.) Segment 11 NE 16th Ave. SW 6th Street Left turn 1.8 2 -> 1 R2, R3, R4, R6, R7, R8 Segment 12 SW 6th Street Coastal lab Right turn 2.1 2 -> 1 R2, R3, R4, R6, R10 R8 Check point 4 End Point: Coastal Eng. Lab (1300 SW 6th Street) *Note: Abbreviations for Anticipated & Potential Scenarios R1Upcoming left/right turn at the immediate/next downstream intersection; R2Current lane is not available downstream (e.g. road incident, work zone or change in channelization of the current lane); a nd R3Passing a stopped-bus at bus stop; R4Giving way to a merging vehicle; R5Gaining speed advantage by overtaking a slower moving vehicle; R6Gaining queue advantage; R7Avoiding a truck/heavy vehicle influence; R8Avoiding the pressure imposed by tailgating; R9Attracted by a better pavement condition;

PAGE 110

1105.1.3 Driving Test Procedure Upon arrival, a check-in procedure was followed to ask each participan t to 1) sign the invehicle experiment informed consent form (AP PENDIX D), 2) complete the background survey form (APPENDIX G), and 3) show their drive rs license to confirm their identity and qualifications for driving. With the informed c onsent form, the participants were fully briefed about the aims of the experiment. During the in-vehicle data collection, each participant was accompanied by the researcher to drive on one of the pre-selected rout es and collect data rela ted to the lane-changing maneuvers. Participants were informed about the driving route and the ty pes of questions they might be asked during driving. Moreover, each dr iver was briefed about the three pre-selected check points during the experiment where they would stop and disc uss their actions. As drivers were proceeding through the developed r oute, the followings were recorded: Potential lane change: the situation in which a lane ch ange could have been attempted, but was not. Attempted (but not successful) lane change: the driver attempted a lane change, but the maneuver was not completed. Completed lane change: the driver completed the maneuver successfully. The space gaps from the completed and attempted lane changes reflect the gap acceptance characteristics for that maneuver. For each of th e maneuvers (potential, attempted and completed) occurring during the test, the time and location information, as well as the corresponding lanechanging scenario, were recorded for furthe r analysis. The scenarios were verified by communicating with the subject. Occasionally, the driver was asked if he/she was considering lane-changing, in order to identify the potential and attempted but not completed maneuvers for such scenario. Table 5-4 presents an example of th e raw data collected for the participant with ID 05-11, including the maneuver types identified.

PAGE 111

111 Table 5-5. In-Vehicle experi ment notes for subject ID: 05-11 Location (including posted spd. limit) Time Scenario Maneuver type Archer Rd, 40mph 4:00 Queue advantage 1 Archer Rd, 40mph 4:03 Queue advantage 3 Archer Rd, 45mph 4:09 Stopped bus 2 Archer Rd, 45 mph 4:15 Overtaking Slow vehicle 1 Archer Rd, 45mph 4:16 Overtaking slow vehicle 3 SW 34th St, 45mph 4:24 Heavy vehicle 2 Newberry Rd, 40mph 4:27 Stopped bus 1 Newberry Rd, 40mph 4:33 Overtaking slow vehicle 2 Newberry Rd, 40mph 4:34 Heavy vehicle 1 SW 34th St, 45mph 4:37 Incoming left turn 3 Newberry Rd, 40mph 4:45 Backup Turning 2 Newberry Rd, 35mph 4:55 Incoming left turn 3 Driver ID: 05-11 Maneuver type = 1 (potential maneuv er), 2 (attempted maneuver), 3 (completed maneuver). Two types of video clips were collected during the driving test. The first is from the four DCs installed in the instrumented vehicle, which capture the traffic on the road, and can be used to generate quantitative field va lues for each maneuver. The other is recorded by a fixed digital camcorder to capture the drivers head/eye movement and the discussions between the investigator and the drivers, so that the driv ers behavior and verb al communication can be retrieved during the data reduction. In the check points, the lane-changing related maneuvers that

PAGE 112

112 occurred during the previous stage were discu ssed. Each driver was asked to clarify their interactions with other vehi cles during the driving test. The in-vehicle data collection experiment was conducted from Sept. 2008 to Jan. 2009. All participants drove on a wee kday afternoon to avoid large tra ffic pattern differences between weekdays and the weekend, or between the AM and the PM. A total of 40 driving tests were conducted; 24 were conducted dur ing the early afternoon traffic on the Newberry route (Figure 5-1), and the remaining 16 were conducted dur ing the PM peak traffic on the Waldo route (Figure 5-2). 5.2 Data Reduction and Analysis Information obtained directly from the in-ve hicle experiment incl ude: 1) participants personal background information; 2) in-vehicle video clips for each driving test; 3) the researchers notes for the lane-changing related maneuvers (potential, at tempted and completed maneuvers) during the driving test. Various analyses were conducted to summarize the information related to the research objectives. Fi rst, each of the three types of maneuvers was identified from the video clips using the researchers notes. Quantitative values for the important factors identified from the focus groups as a ffecting each lane-changing related scenario were then obtained. Next, the statistica l summaries of the lane-changing behavioral variab les related to the subject and surrounding vehicles, such as gaps speeds and accelerations, were presented with the corresponding distributions. Last ly, various cluster analyses were performed (similar to that conducted for the focus group data) to categorize the participating driv ers according to their behavior. The detailed description of each anal ysis procedure, along with a summary of the results, is provided below.

PAGE 113

1135.2.1 Video Data Reduction During the 40 in-vehicle driving tests, a to tal of 601 completed la ne changes and 199 attempted but unsuccessful lane changes were collected. In addition, the researcher found another 205 potential lane-changi ng maneuvers. Table 5-6 presents the number of maneuvers for each driver in the experiment. Table 5-6. Driver-based numb er of maneuvers collected in the In-Vehicle experiment Number of maneuvers ID Potential Attempted Completed 05-01 3 9 15 05-02 5 3 11 05-03 3 4 17 05-04 6 3 9 05-05 4 5 14 05-06 7 4 17 05-07 3 3 14 05-08 5 3 16 05-09 8 4 21 05-10 3 3 10 05-11 4 5 12 05-12 7 2 23 05-13 4 6 10 05-14 3 6 9 05-15 5 4 18 05-16 5 3 15 05-17 4 7 11 05-18 6 4 10 05-19 5 5 14 05-20 7 6 17 05-21 5 7 19 05-22 9 7 25 05-23 4 4 16 05-24 3 8 14 05-25 2 6 9 05-26 6 5 15 05-27 7 8 21 05-28 3 4 13 05-29 8 7 17 05-30 7 6 19 05-31 5 5 16 05-32 6 4 13 05-33 6 4 16 05-34 3 3 8 05-35 8 8 22 05-36 4 3 9 05-37 3 5 16 05-38 6 6 20 05-39 7 4 16 05-40 6 6 14 Total 205 199 601

PAGE 114

114 Using the maneuver time record ed during the driving tests, each lane-changing related maneuver (potential, attempted and completed) was located in the video clip. The following information was obtained from this in-vehicle experiment (related to the important factors identified from the focus groups): 1. Information obtained directly from video clips. This information was observed directly from the video clips for each lane-changing related within the driving test, which includes: Traffic Signal Status: the traffic signal status for the downstream signal intersection; the two alternative states are re d and green (including yellow). Number of Lanes: the number of lanes of the current road segment. Vehicle Type: the type of the vehicles involve d in the lane-changing situation. Subject Vehicle Speed and Location: the speed and the geographical coordinates of the subject vehicle can be observed directly from the video, and captured by the GPS installed in the vehicle. Presence of Pedestrians and Cyclists: whether there are pedestrians and cyclists present in the scenario. Level of Congestion: the number of vehicles in a 600ft vicinity, 300 ft behind and 300 ft in front of the subject vehicle at the tim e, were obtained and used as a surrogate for congestion level. 2. Information interpolated from frame-by-frame images. First, the lane-changing video clips were separated into frame-by-frame images (0 .5 sec). Next, the estimation method proposed in Section 3.2 (using lane width and focal di stance of the digital cameras) was used to approximate the vehicle gaps before and after initiating the lane change. Space gaps, rather than time gaps, were measured to minimize errors, since under congested conditions the speeds of the vehicles do not vary significan tly, and a minimal safe spacing is always maintained by drivers. The change of the spacing was estimated using frame-by-frame analysis to obtain other vehicles speeds (the speed of the subject vehicle was acquired directly from the GPS instrument). Thus, the speed change (accelerat ion/deceleration) was

PAGE 115

115 estimated by considering the time interval between successive frames and the speeds measured from these frames. The information obtained from this pr ocedure includes: Vehicle Gaps: the space gaps between any two vehi cles of interest in the image. Surrounding Vehicles Speeds: the speed of any surrounding vehicles. Acceleration/Deceleration: the speed change of any surrounding vehicles appeared. Lane-Changing Duration: it is measured manually starting as the subject vehicle moves laterally and the edge of its first head light crosses the lane delimiter, and ending as the last taillight crosses th at line (Salvucci and Liu, 2002). 3. Information obtained from Google Earth/Maps by using the necessary information from the video/images. For some of the distances, the starting point is the location of the subject vehicle, which can be acquired from the GPS in strument (geographic coordinates). For some cases, the ending point is out of the images s cale, and consequently the position can only be obtained by referring to landmarks or position/lo cation references along the test routes. By mapping the ending position as a geographic re ference in Google Earth (using the ruler function), the corresponding di stance can be obtained. The distances obtained from this procedure include: Distance to the Downstream Intersection: distance from the current position of the subject vehicle to the downstream intersection. Distance to the Next Bus Stop: distance from the current position of the subject vehicle to the next bus stop. Distance to the Upcoming Right/Left Turn: distance from the current position of the subject vehicle to the upcoming right/left turn. By the end of the data reduction, each lane-cha nging scenario was associated with a list of important factors, with multiple sets of corre sponding field values collected from the invehicle experiment. Three types of datasets : completed lane-changing data, attempted but unsuccessful lane-changing data, a nd potential lane-changing data were obtained. An example of

PAGE 116

116 the lane-changing data for the participant with ID 05-11 is provided in Tables 5-7 (completed maneuvers), 5-8 (attempted but unsuccessful maneuvers), and 5-9 (potential maneuvers). Table 5-7. Data collected from completed lane changes (ID = 0511) LC # Invoking scenarios Important factors & values Lag gap (ft) Lead gap (ft) 1 Upcoming left turn Cgst. = 4 Signal = red Spdl/ts = 20/23, Dist. = 96 ft N/A 34 2 Queue advantage Que diff = 3 vehs, Dist. = 4 blk Cgst = 13, Sig. = red 248 15 3 Overtaking slow vehicle Dist. = 2363 ft, Spdl/ts = 45/35 Cgst. = 15, Mood = relax 53 22 4 Stopped bus Cgst. = 9, Next stop = 1971 ft Dist bus = 60 ft, # waiting = 2 35 19 5 Incoming right turn Dist. = 150 ft, Spdl/ts = 45/30 familiar, Cgst. = 14, Ped = no 41 23 6 Incoming left turn Cgst. = 11 Signal = red Spdl/ts = 35/30, Dist. = 113 ft 30 18 7 Overtaking slow vehicle Dist. = 2594 ft, Spdl/ts = 45/40 Cgst. = 3, Mood = med. 212 233 8 Vehicle merge Cgst. = 8, Spdl/ts = 40/44 Agg. = low, Dist. = 3 blocks VType = sedan/suv 276 219 9 Overtaking slow vehicle Dist. = 2910 ft, Spdl/ts = 45/40 Cgst. = 10, Mood = med. 48 31 10 Incoming left turn Cgst. = med Signal = red Spdl/ts = 40/30, Dist. = 150 ft 110 85 11 Lane channel. change Cgst. = 18, Dist. = 55 ft. Spdl/ts = 40/42 13 17.5 12 Incoming right turn Dist. = 85 ft, Spdl/ts = 45/40 Fam. = Y, Cgst. = 12, Ped = N 27 16 Driver ID: 05-11 Note: Spdl/ts: speed limit and the real travel speed, mph. Cgst: level of traffic conge stion on the target lane. Dist: distance to the incoming turn position, the next bus stop, etc.

PAGE 117

117 Table 5-8. Data collected from attempte d but unsuccessful lane changes (ID = 0511) LC # Invoking scenarios Important factors & values Lag gap (ft) Lead gap (ft) 1 Incoming left turn Cgst. = 3, Signal = red Spdl/ts = 20/18, Dist. = 66 ft N/A 24 2 Overtaking slow vehicle Dist. = 3 blk, Spdl/ts = 45/30 Cgst. = 15, mood = relax 13 7.4 3 Vehicle merge Cgst. = 7, Spdl/ts = 40/28 Agg. = low, dist. = 3 blocks veh type = sedan/suv, left mer. 11.5 14.5 4 Overtaking slow vehicle Dist. = 1 blk, Spdl/ts = 45/30 Cgst. = 13, mood = med. 35 23 5 Incoming left turn Cgst. = 17, Signal = red Spdl/ts = 45/35, Dist. = 526 ft 8.5 9 Driver ID: 05-11 Table 5-9. Data collected from potential lane changes (ID = 0511) LC # Invoking scenarios Important factors & values Lag gap (ft) Lead gap (ft) 1 Heavy vehicle Spdl/ts = 45/35, Uncft = N Cgst. = 19, VType = SUV 14.5 7 2 Backup turning 9 21.5 3 Queue advantage Que diff = 2 vehs, Dist. = 2 blk Cgst = 9, Sig. = red 47 6.5 4 Stopped bus Cgst. = 13, Next stop = 1.5 blk Dist bus = 46 ft, # waiting = 4 31.5 13 Driver ID: 05-11 5.2.2 Distributions of Selected Lane-Changing Variables The lane-changing related quantitative values obtained from the in-vehicle data collection include speeds, accelerations, traffic de nsity, different types of spacing gaps, etc. A sketch of the vehicles involved in lane cha nging and the corresponding variables is presented in Figure 5-3.

PAGE 118

118 Figure 5-3. Vehicles involved in a la ne-changing maneuver a nd related variables Table 5-10 summarizes statistics of the variab les obtained from the completed lane changes related to the subject, fr ont, and lead/lag vehicles. Table 5-10. Statistics of variable s related to completed lane changes Variables Mean Std dev. Median Min. Max. Subject vehicle Sub. spd (mph) 21.4 7.23 23.7 2.6 44.6 Sub. acc (mph/s) -0.14 1.21 0.07 -4.96 3.27 Sub. lane den. 10.4 4.63 10 0 23 Relation with lead & lag vehicles Lead gap (ft) 63.1 55.1 55.3 2.3 295.6 Lag gap (ft) 73.4 46.9 64.2 4.1 221.7 Rela. lead spd. (subject lead) -0.84 4.92 -0.5 -13.8 9.5 Rela. lag spd. (subject lag) 0.55 5.45 0.3 -11.6 10.2 Tar. lane den. 8.93 5.37 10 1 22 Relation with front vehicle Front gap (ft) 69.3 47.2 66.1 6.6 232.1 Rela. front spd. (subject front) -0.23 3.12 -0.9 -9.4 8.4 Note: The traffic densities on the subject and targ et lanes were interpolated with the number of vehicles in a 600 ft vicinity 300 ft behind and 300 ft in front of the subject vehicle, since generally the lane-changing dr ivers only consider the congestion in the vicinity. As some lane changes didnt have lead/lag vehi cle, the lead/lag gaps for these cases were not included in the statistics. As shown in Table 5-9, the accepted lead gaps fo r successful lane changes vary from 2.3 to 295.6

PAGE 119

119 feet, with a mean of 63.1 feet. The accepted lag gaps vary from 4. 1 to 221.7 feet, with a mean of 73.4 feet. The front gaps vary from 6.6 feet to 232 .1 feet, with a mean of 69.3 feet. The relative speeds for lead/lag/front vehicles are defined as the speed of the subject vehicle le ss the speed of the lead/lag/front vehicle. As expected, the mean of the relative lead speeds is positive, and the mean of the relative lag speeds is negative. This indicates that in a lane-changing maneuver, for accepted situations on average, the s ubject vehicle is slower relative to the lead vehicle and faster relative to the lag vehicle. Similarly, the mean de nsity of the target lane is slightly lower than that of the subject lane, which m eans that the drivers are more capab le and/or willing to merge to the less congested lane. The dist ributions of these variables within all completed lane changes are presented in Figure 5-4.

PAGE 120

120 Figure 5-4. Distributions of lane-changing variables for the co mpleted maneuvers (a) Subject speed (b) Subject acceleration (c) Subject lane density (d) Target lane density (e)Lead gap (f) Lag gap (g) Front gap (h) Relative front speed (i) Relative lead speed (j) Relative lag speed

PAGE 121

121 In addition to the completed lane changes, the statistical summaries of the variables obtained from the attempted and potential maneuvers are presented in Tabl e 5-11 and Table 5-12 respectively. As shown in Table 5-11, compared to the completed lane changes, the lead and lag gaps in the attempted lane changes have much smaller means and varian ces. The target lane density is higher. There may be because that th e main reason for the attempted lane-changes to not be completed is that the gaps in the target la ne are too small, which is generally caused by the high density in the target lane. Consequently, su ch maneuvers always have smaller lead and lag gaps, and larger target lane dens ity. The mean subject speed in the attempted maneuvers is also smaller since drivers in these situations may ad just speeds to place the subject vehicle in an appropriate position for merge, and the most frequently used adjustme nt is to slow down slightly. This additionally causes the relative sm aller value of the average front gap. Table 5-11. Statistics of variable s related to attempted lane changes Variables Mean Std dev. Median Min. Max. For subject vehicle Sub. spd (mph) 17.7 3.58 19.6 5.9 32.6 Sub. acc (mph/s) -0.22 1.03 -0.16 -4.58 3.31 Sub. lane den. 11.2 3.83 11 2 19 Relation with lead/lag vehicle Lead gap (ft) 7.62 5.43 7.3 -3.6 25.6 Lag gap (ft) 6.47 3.97 6.2 -4.3 19.7 Rela. lead spd. (subject lead) -5.49 3.05 -5.2 -10.1 6.9 Rela. lag spd. (subject lag) -4.42 3.76 -4.7 -9.7 5.3 Tar. lane den. 15.2 3.03 15 10 26 Relation with front vehicle Front gap (ft) 53.7 40.8 51.6 4.8 249.3 Rela. front spd. (subject front) 1.54 3.83 1.9 -9.4 10.7 Note: The traffic densities on the subject and targ et lanes were interpolated with the number of vehicles in a 600 ft vicinity 300 ft behind and 300 ft in front of the subject vehicle, since generally the lane-changing dr ivers only consider the conge stion in their vicinity. As some lane changes didnt have lead/lag ve hicle, the front gaps for these cases were not included in the statistics.

PAGE 122

122 Table 5-12. Statistics of variable s related to potential lane changes Variables Mean Std dev. Median Min. Max. For subject vehicle Sub. spd (mph) 25.3 10.27 25.7 3.9 42.1 Sub. acc (mph/s) -0.21 1.44 -0.09 -4.47 3.74 Sub. lane den. 7.43 3.19 8 0 21 Relation with lead/lag vehicle Lead gap (ft) 54.3 29.7 49.7 3.5 273.8 Lag gap (ft) 47.6 36.3 44.3 4.5 232.5 Rela. lead spd. (subject lead) -0.75 3.83 -0.6 -9.6 7.6 Rela. lag spd. (subject lag) 0.69 4.92 0.5 -10.3 8.4 Tar. lane den. 11.2 5.96 11 2 18 Relation with front vehicle Front gap (ft) 55.2 59.8 51.1 8.6 281.9 Rela. front spd. (subject front) -0.41 3.92 -0.5 -11.9 10.2 Note: The traffic densities on the subject and targ et lanes were interpolated with the number of vehicles in a 600 ft vicinity 300 ft behind and 300 ft in front of the subject vehicle, since generally the lane-changing dr ivers only consider the conge stion in their vicinity. As some lane changes didnt have lead/lag vehi cle, the lead/lag gaps for these cases were not included in the statistics. The distributions of these vari ables within all attempted lane changes are presented in Figure 5-5. For the potential maneuvers (Tab le 5-12), both the average lead and lag gaps (49.3 ft and 47.3 ft) are smaller than those of the complete d lane changes (63.1 ft and 73.4 ft), which therefore caused the higher targ et lane density values. The aver age front gap is smaller, which may be because of the higher traffic density. The rest of factors, such as the subject speed, acceleration, relative lead/lag sp eed, do not change much. The distributions of the variables within all potential lane changes are presented in Figure 5-6.

PAGE 123

123 Figure 5-5. Distributions of lane-changing variables for the at tempted maneuvers (a) Subject speed (b) Subject acceleration (c) Subject lane density (d) Target lane density (e)Lead gap (f) Lag gap (g) Front gap (h) Relative front speed (i) Relative lead speed (j) Relative lag speed

PAGE 124

124 Figure 5-6. Distributions of lane-changing variables for the potential maneuvers (a) Subject speed (b) Subject acceleration (c) Subject lane density (d) Target lane density (e)Lead gap (f) Lag gap (g) Front gap (h) Relative front speed (i) Relative lead speed (j) Relative lag speed

PAGE 125

1255.2.3 Cluster Analysis for Dr iver Type Classification With the three types of in-vehicle lane-c hanging datasets obtained for each driver (completed, attempted and potential), clustering analysis similar to what was conducted for the focus group study was performed to cl assify the in-vehicle drivers. The objective of this step is to obtain a scheme that can effectively classify drivers into different gr oups. Two classification schemes were designed and conducted. One wa s based on the drivers background information acquired from the participants check-in proced ure. The other used the drivers lane-changing aggressiveness indices measured from the beha viors that occurred during the in-vehicle driving test. These classification results were compared, along with the one obtained from the focus group study. 5.2.3.1 Classification scheme I driver background-based scheme In this scheme, the same driver background in formation (driver aggressiveness) as the one used in the focus group study was used. As pr esented in Table 5-13, the Aggressiveness columns on the left side provi de the self-evaluation and th e perceived friends evaluation obtained from the background survey during the participant prescreening. An overall aggressiveness for each participant was calcula ted by averaging the self-evaluation and the friends evaluation values. The results show that the self-evaluation value is generally slightly less than the friends evaluation (Selfavg.= 5.14, Friendsavg.= 5.53), which was also found to be true in the focus group data.

PAGE 126

126 Table 5-13. Driver-Based likelihood of executing a DLC (In-Vehicle experiment) Aggressiveness (1 10) In-Vehicle Results (c alculated by Eq. 5-1) ID SelfFriends Overall R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 05-01 7-8 7 7.25 4.2 2.8 4.8 4. 5 4.2 4.9 3.5 2.8 3.9 4.0 05-02 5 6 5.5 3.9 3.2 4.5 4. 2 4.0 1.1 2.6 4.2 4.4 3.8 05-03 6 8 7.0 4.4 2.8 4.8 4. 0 4.2 1.1 3.6 2.7 3.9 4.7 05-04 6 5 5.5 4.3 4.2 3.7 4. 4 3.8 1.7 1.9 4.5 4.1 4.6 05-05 4 4-5 4.25 4.2 4.3 3.8 4. 5 3.8 1.7 2.0 4.4 4.2 4.5 05-06 3 3-4 3.25 3.2 3.8 3.9 3. 7 2.9 1.5 3.2 3.9 3.3 3.9 05-07 7 8 7.5 4.3 2.7 4.8 3. 6 4.2 1.3 3.6 2.8 3.9 4.0 05-08 5-6 7 6.25 4.4 2.8 4.7 4. 0 4.1 4.9 3.7 2.7 3.9 3.6 05-09 3 4 3.5 3.2 3.8 3.9 3. 7 2.9 1.5 3.2 3.9 3.1 3.9 05-10 6-7 7 6.75 4.3 2.8 4.8 4. 1 4.2 5.0 3.6 2.7 3.9 3.0 05-11 5-6 6 5.75 3.8 3.2 4.5 4. 0 4.0 1.2 2.6 4.2 4.4 3.8 05-12 3 4 3.5 4.3 4.4 3.8 4. 4 3.8 1.7 2.0 4.5 4.2 4.5 05-13 6 5 5.5 4.2 4.3 3.9 4. 5 3.8 1.8 1.9 4.4 4.3 3.5 05-14 3 3 3.0 4.4 3.8 3.9 3. 7 2.9 1.5 3.2 4.0 3.1 3.9 05-15 4 3-4 3.75 4.3 4.3 3.8 4. 4 3.8 1.7 1.9 4.4 4.2 4.5 05-16 7 8 7.5 4.4 2.9 4.7 4. 5 4.2 5.0 3.6 2.7 3.9 4.0 05-17 6 6-7 6.25 4.3 4.2 3.8 4. 4 3.8 1.7 1.9 4.4 4.3 4.5 05-18 8 8-9 8.25 4.2 2.8 4.8 3. 1 4.2 3.3 3.7 2.8 3.9 3.7 05-19 6 7 6.5 4.3 2.7 4.9 4. 0 4.2 5.0 3.6 2.7 3.9 4.0 05-20 4 5 4.5 3.5 3.8 3.9 3. 7 2.9 3.9 3.2 3.9 3.1 3.9 05-21 4-5 6 5.25 4.2 4.3 3.8 4. 5 3.8 1.7 1.9 4.5 4.2 4.4 05-22 5 5 5.0 4.3 4.4 3.7 4. 4 3.8 1.9 1.9 4.4 4.3 4.5 05-23 7 8 7.5 3.8 3.2 4.5 4. 7 4.0 1.1 2.6 4.3 4.4 3.8 05-24 7 6 6.5 3.8 3.2 4.5 4. 2 4.0 1.2 2.7 4.2 4.4 3.9 05-25 2 3 2.5 3.3 3.8 3.9 3. 7 2.9 1.5 3.2 3.9 3.1 3.9 05-26 5 5-6 5.25 3.8 3.2 4.5 4. 1 4.0 1.1 2.6 4.2 4.4 3.3 05-27 7 6-7 6.75 4.4 2.7 4.8 4. 0 4.2 1.5 3.6 2.7 3.9 4.5 05-28 5 6 5.5 3.7 3.2 4.5 4. 6 4.0 1.1 2.7 4.3 4.4 3.8 05-29 7 7 7.0 3.8 3.2 4.6 3. 4 4.0 1.2 2.6 4.2 4.4 3.7 05-30 6 6-7 6.25 2.9 3.8 3.9 3. 7 2.9 1.5 3.2 3.9 3.1 3.9 05-31 4 3-4 3.75 4.3 4.2 3.8 4. 4 3.8 1.7 1.9 4.4 4.2 4.5 05-32 4-5 4-5 4.5 4.2 4.3 3.7 3.5 3.8 1.8 1.9 4.5 4.3 4.6 05-33 3 3-4 3.25 3.3 3.8 3.9 3. 6 2.9 1.5 3.2 3.9 3.1 2.9 05-34 2 3 2.5 3.2 3.7 3.9 3. 6 2.9 1.5 3.3 3.9 3.2 3.9 05-35 4 4 4.0 3.1 4.3 3.8 4. 4 3.8 1.7 1.9 4.4 4.2 4.5 05-36 5 4-5 4.75 3.7 3.2 4.5 4.7 4.0 5 2.6 4.2 4.4 3.8 05-37 6 6-7 6.25 3.8 3.2 4.4 4. 2 4.0 1.2 2.6 4.3 4.5 3.7 05-38 6 5 5.5 3.9 3.2 4.5 3. 7 4.0 5.0 2.6 4.2 4.4 3.8 05-39 4 5 4.5 4.3 4.4 3.8 4. 4 3.8 1.7 1.9 4.4 4.2 4.5 05-40 5-6 6 5.25 3.1 3.8 3.9 3.7 2.9 5 3.3 3.9 3.1 3.9 Average 5.14 5.54 5.33 3.9 3.6 4.2 4.1 3.7 2.3 2.8 3.9 4.0 4.0 The right side of the table (In-Vehicle Results) presen ts the likelihood of changing lanes on each DLC scenario (R1 through R10) fo r each participant, which was calculated from the lane-changing maneuvers occurring during the in-vehicle test by Eq. (5-1): maneuvers potential of # changes lane attempted of # changes lane completed of # changes) lane attempted of # changes lane completed of (# 5 P (5-1) To obtain the numbers used in this calculation, the in-vehicle field data were first grouped by

PAGE 127

127 driver and lane-changing scenarios. For each driver under a DLC scenario, the number of completed lane changes plus the number of a ttempted lane changes indicates the number of situations accepted by th e particular driver. This number reflects the corresponding level of acceptance for this DLC scenario by the particular driver. The likelihood value P in Eq. (5-1) was then calculated and scaled to five to obtain the same quantitative range (0-5) as used in the focus group study. By following the same classification procedure as in the focus group analysis, the K-means algorithm (provided in APPENDIX I) was first used to cluster the n (n = 40) partic ipants based on overall aggressiveness into k (k = 1, 2, 3, 4, or 5) partitions, k < n. By setting the cluster num ber as 1, 2, 3, 4, and 5 respectively, centr oids for the clusters were obtained as follows: 1) for cluster number = 1, centro id for each cluster is 5.33; 2) for cluster number = 2, centroids for each cluster are 3.78 and 6.35; 3) for cluster number = 3, centroids for each cluster are 3.39, 5.16 and 6.90; 4) for cluster number = 4, centroids for each cluster are 3.30, 4.81, 5.83 and 7.14; 5) for cluster number = 5, centroids for each cluster are 3.22, 4.42, 5.4, 6.55 and 7.60. Next, the scenario-based level of likelihood values, as presented in the right side of Table 5-12, were used to decide the most appropriate cluster number. The overall intra-cluster variance on the level of likelihood for each lane-chang ing situation was calculated, and accumulated across all scenarios (R1 through R 10) by using Eq. (4-2). The overall intra-cluster variance (W) value for each classification was calculated as )1( W177.79, )2( W158.40, )3( W140.92, )4( W133.31 and )5( W129.59. The Hartigan index, which indicates the intra-cluster dissimilarity that will be removed by splitting the k clusters into k+1 clusters, was then calculated to determine the appropr iate number of clusters. By usi ng Eq. (4-3), the indices for k equaling 1, 2, 3 and 4 were calculated as H(1) = 4.65, H(2) = 5.59, H(3) = 2.05 and H(4) = 1.00. A rather small Hartigan index was found to occur in both H(3) (=2.05) an d H(4) (=1.00), which

PAGE 128

128 means by splitting the 3/4 clusters into 4/5, the dissimilarity is not removed as much. Given that the selection of the number of clusters is not quan titatively strict, either of these may be chosen as the appropriate number of clusters. Figure 5-7 provides the results of analysis for the number of clusters ranging from 1 to 5. When the cluste r number is larger than 4, the intra-cluster dissimilarity does not decrease much. However, it is difficult to infer/deduce the exact optimal number from the figure. Furthe r behavior-based clustering anal ysis is conducted to help the classification. Figure 5-7. Clustering analysis results based on driver background 5.2.3.2 Classification scheme II driver behavior based scheme In this scheme, a second classification scheme was proposed based on the driver behavior that occurred during the in-vehi cle driving test. Three quantitativ e measures of driver behavior were selected to evaluate the field driving a ggressiveness of each participant as follows (AAA Foundation, 2009): Number of attempted and completed discretionary lane changes, Number of completed lane changes without turn signal, and Number of improper drivi ng behaviors, including 1) failure to yield right of way, 2) failure to obey traffic signs, and 3) driving too fast for conditions or in excess of posted speed limit.

PAGE 129

129 Table 5-14 presents the number of various ma neuvers that occurred during the in-vehicle driving test, as well as the co rresponding aggressiveness index (A I) values calculated from the measures. An overall field aggres siveness index (FAI) was measured for each of the participants. For the number of attempted and completed DLCs (Attempted and completed DLCs), the AI (AI1) value is calculated by Eq. (5-2): 10* )min()max( )min(1NN NN AIi (5-2) where, AI1: is the aggressiveness inde x for the measurement of number of attempted and completed DLCs, Ni: is the number of attempted and completed DLCs by participant i, min(N): is the minimal number of attempted and completed DLCs for all participants, and max(N): is the maximal number of attempted and completed DLCs for all participants. For the number of completed lane change s without signal ahead (Completed LCs w/o signal ahead), the AI (AI2) value is calculated by Eq. (5-3): 10* )/min()/max( )/min( // / / ,/ 2 total ow total ow total ow i total i owNN NN NN NN AI (5-3) where, AI2: is the aggressiveness index for the measurement of number of completed lane changes without signal ahead, Nw/o: is the number of completed lane changes without signal ahead by participant i, and Ntotal: is the total number of completed lane changes by participant i. For the number of improper driving beha viors (Improper driving behaviors), the AI (AI3) value is calculated by Eq. (5-4): 10* )min()max( )min(, 3 total total total itotalN N N N AI (5-4) where, AI3: is the aggressiveness index for the meas urement of number of improper driving behaviors, Ntotal, i : is the total number of improper driving behaviors by participant i, which is calculated by adding the number of failures to yield right of way (iyieldN,), the number of failures to obey the traffic signs (iobeyN,), and the number of driving too fast for

PAGE 130

130 conditions or in excess of posted speed limit (ispeedyN,), min(Ntotal): is the minimal number of improper driving behaviors, and max(Ntotal): is the maximal number of improper driving behaviors. Table 5-14. Drivers FAI interpolat ed from selected field behaviors Attempted and completed DLCs Completed LCs w/o signal Improper driving behaviors (#) ID N AI1 Nw/o / Ntotal AI2 Nyield NObey NSpeedy AI3 FAI 05-01 19 6.19 8 / 20 7.11 13 5 8 10 7.77 05-02 9 1.43 9 / 16 10 9 4 7 7.39 6.27 05-03 16 4.76 10 / 22 8.08 13 7 5 9.57 7.47 05-04 12 2.86 5 / 14 6.35 6 2 5 4.35 4.52 05-05 14 3.81 5 / 19 4.68 9 6 6 7.83 5.44 05-06 16 4.76 3 / 22 2.42 5 4 5 4.78 3.99 05-07 20 6.67 7 / 19 6.55 8 5 8 7.83 7.02 05-08 22 7.62 9 / 21 7.62 7 6 7 7.39 7.54 05-09 14 3.81 5 / 20 3.42 3 1 3 1.74 2.99 05-10 18 5.71 7 / 15 8.30 10 2 5 6.09 6.70 05-11 12 5.24 6 / 17 6.27 8 5 5 6.52 6.01 05-12 20 6.67 4 / 28 2.54 7 4 7 6.52 5.24 05-13 11 2.38 8 / 15 9.48 5 3 4 3.91 5.26 05-14 10 1.90 5 / 14 6.35 2 1 3 1.30 3.18 05-15 17 5.24 5 / 23 3.86 8 4 5 6.09 5.06 05-16 17 5.24 8 / 20 7.11 13 5 5 8.70 7.02 05-17 16 4.76 6 / 16 6.67 6 5 4 5.22 5.55 05-18 27 10 6 / 15 7.11 9 4 6 6.96 8.02 05-19 17 5.24 9 / 19 8.42 12 3 7 8.26 7.31 05-20 18 5.71 5 / 22 4.04 4 2 2 2.17 3.97 05-21 21 7.14 4 / 24 2.96 8 3 7 6.52 5.54 05-22 13 3.33 9 / 30 5.33 8 7 6 7.83 5.50 05-23 15 4.29 8 / 31 6.77 12 4 6 8.26 6.44 05-24 17 5.24 6 / 19 5.61 11 3 5 6.96 5.94 05-25 12 2.86 3 / 14 3.81 2 0 1 0 2.22 05-26 15 4.29 7 / 20 6.22 8 6 8 8.26 6.26 05-27 24 8.57 7 / 26 4.79 10 5 7 8.26 7.21 05-28 15 4.29 7 / 18 6.91 10 3 6 6.96 6.05 05-29 19 6.19 5 / 22 4.04 13 4 7 9.13 6.45 05-30 20 6.67 1 / 24 0.74 8 2 4 4.78 4.06 05-31 16 4.76 4 / 21 3.39 9 6 7 8.26 5.47 05-32 12 2.86 5 / 18 4.94 7 4 7 6.52 4.77 05-33 15 4.29 2 / 21 1.69 5 2 3 3.04 3.01 05-34 6 0 3 / 13 4.10 7 2 4 4.35 2.82 05-35 25 9.05 0 / 27 0 8 5 7 7.39 5.48 05-36 7 0.48 6 / 14 7.62 10 5 9 9.13 5.74 05-37 16 4.76 6 / 21 5.08 9 4 7 7.39 5.74 05-38 21 7.14 4 / 25 2.84 14 4 7 9.57 6.52 05-39 15 4.29 3 / 21 2.54 10 3 9 8.26 5.03 05-40 15 4.29 1 / 19 0.94 4 3 5 3.91 3.05 Average 16.10 4.87 5.5 / 19.9 5.17 8.25 3.83 5.70 6.44 5.49 With these calculations, the overall field aggre ssiveness index (FAI) was then calculated as the average of the three indices AI1, AI2, and AI3, which was used for the driver clustering analysis that followed.

PAGE 131

131 By following the same classification procedure as in scheme 1, the K-means algorithm (as provided in APPENDIX I) was first used to cluster the n (n = 40) partic ip ants based on FAI into k (k = 1, 2, 3, 4, or 5) partitions, k < n. By setting the clus te r number as 1, 2, 3, 4, and 5 respectively, centroids for the clusters were obtained as follows: 1) for cluster number = 1, centro id for each cluster is 5.49; 2) for cluster number = 2, centroids for each cluster are 3.51 and 6.24; 3) for cluster number = 3, centroids for each cluster are 3.25, 5.43, and 7.0; 4) for cluster number = 4, centroids for each cluster are 3.25, 5.24, 6.19 and 7.42; 5) for cluster number = 5, centroids for each cluster are 3.25, 4.98, 5.60, 6.34 and 7.42. Next, the reason-based level of likelihood valu es, as presented in the right columns of Table 5-12, were used to decide the most appropriate cluster nu mber. The overall intra-cluster variance on the level of likelihood for each la ne-changing situation was calculated, and accumulated across all reasons using Eq. (4 -2). The overall intra-cluster variance (W) value for each classification was calculated as: )1(W177.79, )2(W155.63 )3(W106.36, )4(W83.91 and )5(W79.85. The Hartigan index, which indicates the intra-cluster dissimilarity that will be removed by splitting the k clusters into k+1 clusters, was calculated to determine the appropriate number of clusters. By us ing Eq. (4-3), the indices for k equaling 1, 2, 3 and 4 were calculated as H(1) = 5.41, H(2) = 17.14, H(3) = 9.63 and H(4) = 1.78. A rather small Hartigan index was found to occur in H(4), which means by splitting the 4 clusters into the 5, the dissimilarity is not removed as much. Fi gure 5-8 provides the results of analysis for the number of clusters ranging from 1 to 5. When the cluster number is larger than 4, the intracluster dissimilarity does not decr ease much. Therefore, it is recommended that the appropriate number of clusters for this scheme is 4. As we can see, this number is also the result obtained from the focus group study, and is rather close to the number of cl usters acquired from background-based classification. Additional co mparison between the two classifications

PAGE 132

132 (background-based and behavior-based) is pr ovided in the following section. Figure 5-8. Clustering analysis results based on driver behavior 5.2.3.3 Results comparison In this section, results from the two classi fications (background-base d and behavior-based) are compared and analyzed. Given that the num ber of clusters obtained from the backgroundbased analysis is 3 or 4, and the cluster number fr om the behavior-based analysis is 4, the ranges for four groups under the two classification sc hemes are obtained. For the background-based classification, the overall aggressiveness is used to categorize the particip ants into four groups defined as L1 (<= 4.0), L2 (4.1 5.3), L3 (5.4 6.4) and L4 (>= 6.5). For the behavior-based one, the FAI is used to categorize the participants into four groups defined as L1 (<= 4.2), L2 (4.3 5.7), L3 (5.8 6.8) and L4 (>= 6.9). Si nce the overall aggressiveness is also used to categorize the focus group drivers, it was found that the range values for each background-based group were slightly lower than t hose obtained from the focus groups (<= 4.1, 4.2 5.6, 5.7 6.5 and >= 6.6). This may be because of the inner di screpancy between the participants in the two experiments. By going through the participants overall aggressiveness values, it was found the average overall aggressiveness value is 5.65 for th e focus group participants and 5.33 for the invehicle participants, which means the focus gr oup drivers are slightly more aggressive.

PAGE 133

133 By applying the clustering ranges from both classifications (background-based and behavior-based) on the 40 particip ants, each of the drivers is ta gged with a corre sponding driver type (A for group L1, B for group L2, C for group L3 or D for group L4) as presented in Table 515. It was found that for more than half of the drivers (23/ 40 = 57.5%), the driver type information obtained from the background-based clas sification is consistent with that from the behavior-based classification. Table 5-15. Consistency between the background -based and behavior-based classifications Background-based Behavior-based ID Overall Agg. Driver Group FAI Driver Group Consistent or not 05-01 7.25 D 7.77 D Y 05-02 5.5 C 6.27 C Y 05-03 7.0 D 7.47 D Y 05-04 5.5 C 4.52 B N 05-05 4.25 B 5.44 B Y 05-06 3.25 A 3.99 A Y 05-07 7.5 D 7.02 D Y 05-08 6.25 C 7.54 D N 05-09 3.5 A 2.99 A Y 05-10 6.75 D 6.70 C N 05-11 5.75 C 6.01 C Y 05-12 3.5 A 5.24 B N 05-13 5.5 C 5.26 B N 05-14 3.0 A 3.18 A Y 05-15 3.75 A 5.06 B N 05-16 7.5 D 7.02 D Y 05-17 6.25 C 5.55 B N 05-18 8.25 D 8.02 D Y 05-19 6.5 D 7.31 D Y 05-20 4.5 B 3.97 A N 05-21 5.25 B 5.54 B Y 05-22 5.0 B 5.50 B Y 05-23 7.5 D 6.44 C N 05-24 6.5 D 5.94 C N 05-25 2.5 A 2.22 A Y 05-26 5.25 B 6.26 C N 05-27 6.75 D 7.21 D Y 05-28 5.5 C 6.05 C Y 05-29 7.0 D 6.45 C N 05-30 6.25 C 4.06 A N 05-31 3.75 A 5.47 B N 05-32 4.5 B 4.77 B Y 05-33 3.25 A 3.01 A Y 05-34 2.5 A 2.82 A Y 05-35 4.0 A 5.48 B N 05-36 4.75 B 5.74 C N 05-37 6.25 C 5.74 C Y 05-38 5.5 C 6.52 C Y 05-39 4.5 B 5.03 B Y 05-40 5.25 B 3.05 A N % of Consistency 57.5%

PAGE 134

134 Further investigation was conducte d to obtain the statistical distribution of each driver to the corresponding driver types by the two classifications, as shown in Table 5-16. From the table, in addition to the drivers whose types are consiste nt within two classifica tions (# = 23), another 16 drivers were found to be tagged with the adjacent groups, which account for 40% of the total number of drivers. This may be explained as that many of these drivers have the overall aggressiveness or FAI values ra ther close to the boundary of gr oup ranges, which may result in being categorized into the adjacent group(s) inst ead of remaining in the same group. Only one driver (05-30) is tagged as type C in background-based classificat ion, while as type A in the behavior-based classification. The quantitative overall aggressiveness and the FAI values (6.25 and 4.06) from Table 5-15 i ndicate that the real/actual difference is not so large. By referring to the in-vehicle video clips, it was found that th e driving occurred on a Friday PM peak, and the field-collected driving behavior is much le ss aggressive because of the heavy traffic. Consequently, the conclusion was drawn that fo r urban lane-changing behaviors, the field driving maneuvers can be somewhat reflected by the background surv ey results, although discrepancies do exist. Table 5-16. Statistical distribution of drivers by the two classifications Behavior Background Type A Type B Type C Type D Type A 6 (15%) 4 (10%) N/A N/A Type B 2 (5%) 5 (12.5%) 2 (5%) N/A Type C 1 (2.5%) 3 (7.5%) 5 (12.5%) 1 (2.5%) Type D N/A N/A 4 (10%) 7 (17.5%) The comparison and analysis further confir med that the in-vehicle drivers can be generalized into four groups, as recommended in the focus group study. Compared to the driver classification scheme used in the focus group st udy, which is fully dependent on the drivers

PAGE 135

135 perceived experience data, the background-based classification is based on both background and field data, while the behavior-based classifi cation depends entirely on the field data. As presented, the results from these three differe nt classification schemes show similarities by connecting the driver characteristics (overall aggressiveness) to the lane-changing maneuvers (FAI). The behavior-based classification and the corresponding ranges we re selected to be applied to further lane-changing model development. As a result, the in-vehicle maneuver data can be classified and used to model the la ne-changing behaviors with different driver characteristics. 5.3 Summary and Conclusions In this chapter, the design and implementa tion processes of the in-vehicle experiment were presented. The results were analyzed to verify the lane-changing process as documented during the focus group study. Two classification schemes were proposed to cluster the invehicle drivers into different groups based on th e drivers background and the driving behavior measured during the data collection. Results from this chapter, including the in-vehicle field data and the selected classificat ion scheme, are to be used for further model development. More specifically, the findings from this experiment are: The quantitative values for the important factors were obtained from the various (completed, attempted and potential) maneuvers that occurred during the driving tests, which are used to develop the scenario-based lane-changing probability model and the gap acceptance model in Chapter 6. With the lane-changing likelihoods for the DLCs calculated from the in-vehicle field data, both the drivers background information and the measurem ents of driving behavior were used to group the drivers into diffe rent types. Two classification schemes, background-based and behavior-based, were conducted. The number of clusters obtained from the background-based analysis is 3 or 4, while the cluster number from the behaviorbased analysis is 4, which is also the number obtained from the focus group study. Comparison between the two classifications i ndicates that the major results of driver classification are consistent except the differe nces induced by some particular immediate field situations. The result of grouping the par ticipating drivers into four types is to be

PAGE 136

136 further used in the scenario-based pr obability model and gap acceptance model development. By this method, the in-vehicl e lane-changing data can be categorized by driver groups, so that the lane-changing beha viors for different type of drivers can be modeled. The in-vehicle experiment is rather helpfu l in validating and conf irming the conclusions from the focus group study, and in collec ting the parameters for lane-changing model development and implementation. However, as disc ussed, one of the issues is that this type of data collection might have some bias in that drivers are likely to modify their behavior when they know they are being observed. Consequently, simulation and calibration endeavors are included to address this dr iver/human-related problem in the following chapters.

PAGE 137

137 CHAPTER 6 MODEL DEVELOPMENT As discussed in Chapter 1, the lane-changing pr ocess is generally m odeled as a sequence of four decision-making steps: lane -changing decision for particular s cenario, target lane selection, gap acceptance and vehicle movement to target lane. This chapter is focused on the modeling components for the probabilistic decision under various DLC scenarios and the gap acceptance procedure. Driver characteristics and field da ta obtained from the i n-vehicle lane-changing maneuvers (potential, attempted and completed) were used to develop the two components. A hierarchical modeling fram ework for the strategies in choice of plan (decision to change lanes) and choice of action (gap acceptance), with the incorporation of driver characteristics, is presented in Figure 6-1. Figure 6-1. Modeling framework for choices of plan and acti on in lane-changing behavior First, in order to model the lane-changing probability under each particular scenario, the combination of completed lane changes and attempted maneuvers are deemed as accept response, while potential maneuvers are recognize d as not accept response. Next, during the gap acceptance modeling, only the attempted a nd completed maneuvers are considered. The completed lane changes are labeled as acceptabl e response, while the attempted maneuvers are

PAGE 138

138 recognized as unacceptable response. By this me thod, the two processes form a special nested logit model (Peot and Smith, 1992; Bhat, 1997; Carrasco and Ortuzar, 2002), even though the target lane selection should be considered in between. The chapter is organized as follows: the de velopment of the lane-changing probability component for the DLC scenarios is presented in Sec tion 6.1. In Section 6.2, with analysis of the gap acceptance characteristics, along with vehicle interactions that occurred in the in-vehicle experiment, the lane changes on urban arterials were classified into three modes: (i) free, (ii) forced, and (iii) competitive/cooperative. A new gap acceptance algorithm is proposed to distinguish and model ea ch of these three modes, respectiv ely. The chapter concludes with a summary of the newly developed lane-changi ng probability and gap acceptance models. 6.1 Scenario-Based Lane-Changing Probability Model In this section, the development of the proba bility function for mode ling the lane-changing decision under each DLC scenario is presented. First, the field lane-changing maneuvers were grouped by lane-changing scenario. Next, different maneuvers (the combination of completed and attempted maneuvers versus potential mane uvers) were identified. With the provided outcome (accept for completed and attempted ma neuvers, not accept for potential ones) for each lane-changing behavior, the probability of changing lanes under each DLC scenario was estimated as a function of the associated impor tant factors and driver types. The following subsections discuss the details of the modeling procedure. 6.1.1 Dataset Overview As presented in Table 6-1, the driver beha vior-based FAI was used to categorize the participants into four groups. Th e number of drivers for types A, B, C and D are 9, 12, 11 and 8, respectively.

PAGE 139

139 Table 6-1. Classification of driver groups for the in-vehicl e experiment (based on FAI) Driver Group ID FAI (0-10) 05-25 2.22 05-34 2.82 05-09 2.99 05-33 3.01 05-40 3.05 05-14 3.18 05-20 3.97 05-06 3.99 Type A-L1 (number of drivers = 9) 05-30 4.06 05-04 4.52 05-32 4.77 05-39 5.03 05-15 5.06 05-12 5.24 05-13 5.26 05-05 5.44 05-31 5.47 05-35 5.48 05-22 5.5 05-21 5.54 Type B-L2 (number of drivers = 12) 05-17 5.55 05-36 5.74 05-37 5.74 05-24 5.94 05-11 6.01 05-28 6.05 05-26 6.26 05-02 6.27 05-23 6.44 05-29 6.45 05-38 6.52 Type C-L3 (number of drivers =11) 05-10 6.7 05-07 7.02 05-16 7.02 05-27 7.21 05-19 7.31 05-03 7.47 05-08 7.54 05-01 7.77 Type D-L4 (number of drivers = 8) 05-18 8.02 Note: The field aggressiveness index (FAI) scales for various driver types were obtained as Type A (<= 4.2), Type B (4.3-5.7), Type C (5.8-6.8) and Type D (>= 6.9).

PAGE 140

140 There are a total of 601 completed lane ch anges, 199 attempted lane changes and 205 potential maneuvers occurring during the in-ve hicle data collection. These maneuvers are grouped by scenario with the number of maneuvers for each DLC scenario as shown in Table 6-2. For each maneuver, in addition to the subject driver type, which can be obtained from Table 6-2, the associated important factors were identified from the focus group study as shown in Table 49. Table 6-2. Number of lane ch anges collected for each scenario LC Reasons Actions Number of Maneuvers Potential 15 Attempted 13 R1 (Stopped bus) Completed 45 Potential 23 Attempted 14 R2 (Vehicle merge) Completed 47 Potential 36 Attempted 41 R3 (Slow vehicle) Completed 167 Potential 15 Attempted 20 R4 (Queue advantage) Completed 49 Potential 12 Attempted 9 R5 (Heavy vehicle) Completed 27 Potential 20 Attempted 0 R6 (Tailgating) Completed 10 Potential 28 Attempted 11 R7 (Pavement) Completed 29 Potential 17 Attempted 16 R8 (Backup turning) Completed 43 Potential 13 Attempted 11 R9 (Pedestrian/scooter) Completed 30 Potential 15 Attempted 20 R10 (Erratic driver) Completed 47

PAGE 141

1416.1.2 Lane-Changing Probability Function Estimation In studying the maneuvers under each of th e lane-changing scenarios, the dependent variable is the outcome of a binary choice (: accept for completed and attempted cases; : not accpet for potential ones). A logistic regression w ith a binary dependent variable was chosen to estimate the probability of changing lanes unde r each DLC scenario as a function of the associated important factors a nd driver types (Ben-Akiva, 19 73; Ben-Akiva and Lerman, 1985; Ben-Akiva and Bierlaire, 2003). Th e logistic regression is essentially a generalized linear model with special advantages for binomial regression (Hosmer and Lemeshow, 2000). The advantages of using the logistic regression instead of the ordinary linear re gress ion in this research are listed below (Albert and Anderson, 1984): If a linear regression is used, the predicted valu es may become greater than 1 or less than 0 if any of the independent variables were moved far enough on the X-axis. Such values are theoretically inadmissible. One of the assumptions of regression is that the variance of Y is constant across values of X (homoscedasticity). This cannot be the case with a binary variable, because the variance is P*Q (P, the proportion of 1s; Q, the proportion of 0s). As P approaches 1 or 0, the variance approaches 0. The significance testing of the coefficients rests upon the assumption that errors of prediction (Y-Y) are normally distributed. Beca use Y only takes the values 0 and 1, this assumption is hard to justify, even approximately. Therefore, the tests of the regression coefficients are suspect if the linear regression were used with a binary dependent variable. By using the logistic regre ssion approach (Nakanishi and Cooper, 1974), the probability function of changing lanes under each scenario is calculated as: )( )(1 )(LCV LCVe e LCP (6-1) where, )( LCP: is the probability of changing lanes under given scenario, and )( LCV: is the utility of changing lanes under a given scenario, which is generally formulated as XT*0 X is the independent variable vector, 0 is the constant and T are the corresponding coefficients.

PAGE 142

142 To assess the impact of important factors on each of the corresponding DLCs, and formulate the lane-changing utilit y function, the maneuvers related to each of the pre-selected DLC scenarios were analyzed as described below. Table 6-3 presents the number of maneuvers for different driv er types that occurred during the stopped-bus scenario. The im portant factors identified for this scenario from focus group study (see Table 4-9) are: factor 1: Traffic congestion in the target lane (Cgst); factor 2: Queue ahead (Que); factor 3: Location of the next downstream stop (LocStop, mile); factor 4: Distance to th e bus (Dist, feet); and factor 5: Number of pers ons at the bus-stop (NPson). Table 6-3. Number of LCs for different driver types dur ing Stopped-Bus scenario Maneuvers Type A(L1) Type B( L2) Type C(L3) Type D(L4) Potential 5 5 3 2 Attempted 2 6 2 3 Completed 7 21 9 8 Total 14 32 14 13 Consequently, a utility function of changing lanes for this scenario is developed as follows: DrvTypeC DrvTypeB DrvTypeA NPson Dist LocStop Que Cgst LCV ** ** )(3 2 1 5 4 3 2 10 (6-2) In this instance, three dummy variables were created as DrvTypeA DrvTypeB and DrvTypeC which are used in regression analysis to represent different driver gr oups (Types A, B and C) included in this study. If a subject driver belongs to type A, then DrvTypeA would be equal to 1, and DrvTypeB and DrvTypeC would be equal to 0. If a subj ect driver is in type D, all three dummy variables would be e qual to 0. In fact, each dummy variable acts as a switch that turns the corresponding driver type parameter on and off in the equation, so that a single regression function can be used to represent multiple driver groups (Kinnear et al., 1974; Khattak

PAGE 143

143 et al., 1996). By this, it is not necessary to de velop separate models fo r each driver group (Golob and McNally, 1997; Brownstone et al., 2000). With the field values obtained from 73 ma neuvers (as shown in Table 6-3), the binomial logistic regression tools in SPSS were used to capture the re lationship between these factors and the lane-changing probability ( Norusis, 2005; Allison, 1999). The estimated results are presented in Table 6 -4 as follows. Table 6-4. Estimated coefficients for the factors in Stopped Bus scenario Factors Coefficients T Value Constant 6.480 3.902 Traffic congestion in the target lane (Cgst) -0.236 -4.517 Queue ahead (Que) 1.218 0.652 Location of the next stop (LocStop) -19.116 -2.778 Distance to the bus (Dist) -0.381 -2.703 Number of persons at the bus-stop (NPson) 0.227 0.217 Driver Type A (DrvTypeA) -2.533 -1.935 Driver Type B (DrvTypeB) -1.303 -1.736 Driver Type C (DrvTypeC) -1.139 -1.807 Among these parameters, the factor of Traffic congestion in the target lane (Cgst) made the most significant contribution to the probability of changing lanes, both in terms of relative magnitude and statistical significance. The factor captures the impact of traffic conditions on the target lane. The negative sign means when the ta rget lane is congested a vehicle has a lower probability to change lanes to avoid the stopped bu s. In addition, the estimated results indicate that Location of the next bus stop (LocStop) and Distance to the bus (Dist) are significant, and both factors have a negative coefficient. These make sense, since the field data were collected from urban arterials wit hout pullouts. As the subject vehi cle approaches the bus, or the bus approaches the next bus stop, it would become more willing to change lanes. As presented in Table 6-4, the driver types affect the lane-c hanging probability differently, and the defensive drivers (type A) tend to have a low probability to change lanes. Two othe r factors: Queue ahead

PAGE 144

144 (Que) and Number of persons at the bus stop (NPson) are not si gnificant at a 90% confidence, although the estimated coefficients (1.218 and 0. 227) seem feasible (with the increasing number of queue length and persons at the bus stop, th e probability of changing lanes increases). By excluding the factors which are not significant at the 90% confiden ce, the explanatory variables for this scenario were selected as: Traffic conges tion in the target lane (Cgst), Location of the next stop (LocStop), Distance to the bus (Dist), and the four driver types. Consequently, the utility function of changing lanes for this scenario is estimated as: DrvTypeC DrvTypeB DrvTypeA Dist LocStop Cgst LCV *139.1 *303.1 *533.2 *381.0 *116.19*236.048.6)( (6-2) Table 6-5 presents the number of maneuvers for different driv er types that occurred during the vehicle merge scenario. The important factor s identified for this scenario (see Table 4-9) are: factor 1: Traffic congestion on the target lane (Cgst); factor 2: Travel speed, and the difference b/t travel speed and speed limit (Spd1, Spd1-Spd2); factor 3: Aggressiveness of the merge (Agg); factor 4: Distance to the next turn (Dist); and factor 5: Merger and the subjec t vehicle type (VehT1, VehT2). Table 6-5. Number of LCs for different driver types dur ing Vehicle Merge scenario Maneuvers Type A(L1) Type B( L2) Type C(L3) Type D(L4) Potential 5 4 6 8 Attempted 3 5 3 3 Completed 13 19 8 7 Total 21 28 17 18 Since the subject vehicle is always the instrumented vehicle (Honda Pilot), the effect of Subject vehicle type (VehT2) was not cap tured in this experiment. Conse quently, the utility function of changing lanes for this scenario is developed as follows: DrvTypeC DrvTypeB DrvTypeA VehT Dist Agg SpdSpd Spd Cgst LCV 1* *)21(*1* )(3 2 1 6 5 4 3 2 10 (6-3)

PAGE 145

145 With the field values obtained from 84 mane uvers (as shown in Table 6-5), the regression tool in SPSS was used to capture the relations hip between these factors and the lane-changing probability. The results are presented in Table 6-6 as follows. Table 6-6. Estimated coefficients for th e factors in Vehicle Merge scenario Factors Coefficients T Value Constant 4.398 4.49 Traffic congestion in the target lane (Cgst) -0.397 -1.89 Subject travel speed (Spd1) 0.422 0.56 Diff. b/t travel speed and spee d limit (Spd1-Spd2) -0.285 -1.99 Aggressiveness of the merge (Agg) -0.693 -0.28 Distance to the next turn (Dist) -0.011 -2.17 Merger vehicle type (VehT1) 0:car, 1: others 1.090 2.02 Driver Type A (DrvTypeA) -1.091 -2.16 Driver Type B (DrvTypeB) 1.711 2.33 Driver Type C (DrvTypeC) 3.166 2.49 Among these parameters, the factor of Diff. b/t travel speed and speed limit (Spd1-Spd2) made the most significant cont ribution to the probability of changing lanes. The negative sign means when a vehicle is at a higher speed than the posted speed, it may have a lower probability to change lanes to give way to the merge vehicle. In addition, th e estimated results indicate that Traffic congestion in the target lane (Cgst), Dista nce to the next turn (Dist) and Merger vehicle type (VehT) are significan t, which reflect the impact of drive environment on the driver. As presented in Table 5-13, the driver type parameters affect the lane-changing probability differently. The median aggressive drivers (types B and C) tend to have a larger probability to change lanes, which is consistent with the findings from the focus group study. The least and most aggressive drivers (types A and D) may choose to slow down to give way or accelerate to prohibit the merge. Two other f actors, Subject travel speed (S pd1) and Aggressiveness of the merge (Agg) are not significant at 90% confid ence. By excluding these two parameters, the explanatory variables for this scen ario were selected as: Traffic congestion in the target lane

PAGE 146

146 (Cgst), Diff. b/t travel speed and speed limit (Spd1-Spd2), Dista nce to the next turn (Dist) and Merger vehicle type (VehT). Consequently, the utility function of changing lanes for this scenario is estimated as: DrvTypeC DrvTypeB DrvTypeA VehT Dist SpdSpd Cgst LCV *166.3 *711.1 *091.1 1*09.1*018.0)21(*285.0*397.0398.4)( (6-4) Table 6-7 presents the number of maneuvers for different driv er types that occurred during the slow vehicle scenario. The important factors id entified for this scenario (see Table 4-9) are: factor 1: Distance to the next turn (Dist); factor 2: Travel speed, and the difference be tween travel speed a nd speed limit (Spd1, Spd1Spd2); factor 3: Congestion on the target lane (Cgst); and factor 4: Drivers mood, hurry or not (Mood) Note: The factor 4 (Mood) is excluded from the estimation, since the va lue of the factor is difficult to be collected from the in-vehicle te st, and is almost impossible to be modeled. Table 6-7. Number of LCs for different driver types dur ing Slow Vehicle scenario Maneuvers Type A(L1) Type B( L2) Type C(L3) Type D(L4) Potential 14 15 5 2 Attempted 8 13 11 9 Completed 39 53 41 34 Total 61 81 57 45 Consequently, a utility function of changing lanes for this scenario is developed as follows: DrvTypeC DrvTypeB DrvTypeA Cgst SpdSpd Spd Dist LCV *)21(*1* )(3 2 1 4 3 2 10 (6-5) With the field values obtained from 244 maneuve rs (as shown in Table 6-7), the regression tool in SPSS was run to capture the relationship between these fa ctors and the lane-changing probability. The results are presented in Table 6-8. Among these parameters, the factor s Distance to the next turn (Dist), Traffic congestion in the target lane (Cgst), Subject speed (Spd1) and Diff. between travel speed and speed limit (Spd1-Spd2) are significant, which reflect the traffic dynamics surrounding the subject

PAGE 147

147 driver. As presented in Table 6-8, the driver types affect the lane -changing probability differently. The most aggressive drivers (type D) tend to have a larger probability to change lanes, which is consistent with the findings from the focus groups The only conflict is that that driver type B tends to have higher intention to change lanes than driver type C, while according to the classification, drivers in type B are characteri zed as less aggressive than those in type C. Considering the coefficien ts for the two factors ( DrvTypeB and DrvTypeC ) are rather close to each other (-0.376 and -0.389), the two variables ma y be grouped together as one united variable with an averaged coefficient value. Consequent ly, the explanatory variables for this scenario were selected as: Distance to the next turn (Dist), Traffic congestion in the target lane (Cgst), Subject travel speed (Spd1) and Diff. between trav el speed and speed limit (Spd1Spd2). The utility function of changing lane s for this scenario is estimated as: ) (*382.0 *703.0 )21(*155.01*037.0*281.0*031.0743.3)( DrvTypeC DrvTypeB DrvTypeA SpdSpd Spd Cgst Dist LCV (6-6) Table 6-8. Estimated coefficients for the factors in Slow Vehicle scenario Factors Coefficients T Value Constant 3.743 3.26 Distance to the next turn (Dist) -0.031 -2.57 Traffic congestion in the target lane (Cgst) -0.281 -3.46 Subject travel speed (Spd1) 0.037 1.66 Diff. b/t travel speed and spee d limit (Spd1-Spd2) -0.155 -2.89 Driver Type A (DrvTypeA) -0.703 -3.24 Driver Type B (DrvTypeB) -0.376 -6.23 Driver Type C (DrvTypeC) -0.389 -2.91 Table 6-9 presents the number of maneuvers for different driv er types that occurred during the queue advantage scenario. The important factors identified for this scenario (see Table 4-9) are: factor 1: Queue length difference (QueDiff); factor 2: Distance to the next turn (Dist); factor 3: Congestion on the target lane (Cgst); and

PAGE 148

148 factor 4: Current signal st atus/green time (CurSig). Consequently, a utility function of changing lanes for this scenario is developed as follows: DrvTypeC DrvTypeB DrvTypeA CurSig Cgst Dist QueDiff LCV * ** )(3 2 1 4 3 2 10 (6-7) Table 6-9. Number of LCs fo r different driver types durin g Queue Advantage scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 5 5 3 2 Attempted 3 7 4 6 Completed 11 22 9 7 Total 19 34 16 15 With the field values obtained from 84 maneuvers (as shown in Table 6-9), the regression tool in SPSS was used to capture the relationship between these factors a nd the lane-changing probability. The results are presented in Table 6-10. Table 6-10. Estimated coefficients for th e factors in Queue Advantage scenario Factors Coefficients T Value Constant 2.503 7.41 Queue length difference (QueDiff) 0.254 9.39 Distance to the next turn (Dist) -0.013 3.74 Congestion on the target lane (Cgst) -0.085 -0.29 Current signal status/red time (CurSig) red:1, not:0 0.492 2.13 Driver Type A (DrvTypeA) -1.694 -4.28 Driver Type B (DrvTypeB) -0.703 -3.93 Driver Type C (DrvTypeC) -0.277 -2.79 Among these parameters, the factor of Queue length difference (QueDiff) made the most significant contribution to the probability of changing lane s. The positive sign means if the queue length difference is larger, the proba bility of changing lane s increases. Estimation results indicate that Distance to the next tu rn (Dist) and Current signal status/red time (CurSig) are also significant, wh ich reflect the concerns of the trade-off between the destination and time spent at the signal. The factor of C urSig has a positive coefficient, which means

PAGE 149

149 drivers are more likely to change lane under red signal. This seems cont radict to our general driving experience. However, c onsidering that most of the driving tests occurred in PM congested traffic, all vehicles would slow down when appro aching a red signal. This may facilitate lane changes since the drivers have more time to observe traffic in the target lane and make appropriate lane changes. As presented in Table 6-10, all driver types affect the lanechanging probability significantly. Aggressive dr ivers (Type D) tend to have a larger probability to change lanes, which is cons istent with the findings from the focus groups. Compared to the other types of driver, driver type A has a much lower probability to change lanes. Another factor, Congestion on the target lane (Cgst), is not sign ificant at 90% confidence. This may be due to the fact that this factor is highly correlated wi th the Queue length difference (QueDiff) factor, and the effect was somewhat captured by the QueDi ff. Drivers tend to cons ider the queue length in front instead of traffic congestion around under PM congested traffic. By excluding the congestion on the target lane (Cgs t), the explanatory variables fo r this scenario were selected as: Queue length difference (QueDi ff), Distance to the next turn (Dist), and Current signal status/red time (CurSig). Conse quently, the utility func tion of changing lanes fo r this scenario is estimated as: DrvTypeC DrvTypeB DrvTypeA CurSig Dist QueDiff LCV *277.0 *703.0 *694.1 *492.0*013.0 *254.0503.2)( (6-8) Table 6-11 presents the number of maneuvers for different driver types that occurred during the heavy vehicle scenario. The important factors identified for this scenario (see Table 4-9) are: factor 1: Travel speed, and the difference b/t travel speed and speed limit (Spd1, Spd1-Spd2); factor 2: Congestion on all lanes (Cgst); factor 3: Personal uncomfortab le with HV (PerUcft); and factor 4: Subject ve hicle type (VehT).

PAGE 150

150 Note: The factor 3 (PerUcft) is excluded from the estimation, since the value of factor is difficult to be collected from the in-vehicle test and is almost impossi ble to be modeled. Since the subject vehicle is always the instrumented vehicle (Honda Pilot), the effect of Subject vehicle type (VehT) was not captured in this experiment. Consequently, a utility function of changing lanes for this scenario is developed as follows: DrvTypeC DrvTypeB DrvTypeA Cgst SpdSpd Spd LCV *)21(*1* )(3 2 1 3 2 10 (6-9) Table 6-11. Number of LCs for different driver types dur ing Heavy Vehicle scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 5 4 2 1 Attempted 2 4 1 2 Completed 5 11 6 5 Total 12 19 9 8 With the field values obtained from 48 maneuvers (as shown in Table 6-11), the regression tool in SPSS was used to capture the relationship betw een these factors and th e probability value. The results are presented in Table 6-12. Table 6-12. Estimated coefficients for the factors in Heavy Vehicle scenario Factors Coefficients T Value Constant 3.314 13.41 Subject travel speed (Spd1) -0.003 -0.36 Diff. b/t travel speed and spee d limit (Spd1-Spd2) -0.065 -1.19 Congestion on the target lane (Cgst) -0.214 -7.39 Driver Type A (DrvTypeA) -1.946 -3.17 Driver Type B (DrvTypeB) -0.781 -2.39 Driver Type C (DrvTypeC) 0.008 -0.83 Among these parameters, only the factors of Congestion on the ta rget lane (Cgst) contributed to the probability of changing lanes. The negative si gn means drivers are less likely to change lanes to avoid followi ng an HV if the target lane is congested. Results from this estimation indicate that the speed-r elated factors, such as Subject travel speed (Spd1) and Diff.

PAGE 151

151 b/t travel speed and speed limit (S pd1-Spd2) are not significant. Th is may be due to the factor that compared to Cgst, driv ers under this situation may not care too much about speed and speed difference. In addition, the speed and spee d difference may be highly correlated with the Cgst factor, and the effects were captured by Cgst. As presented in Table 6-12, the driver types affect the lane-changing proba bility differently. The driver types C tends to have similar effect as the driver type D, which may be grouped together as one united variable, while the driver types A and B are not will ing to change lanes compared to the driver types C and D. By excluding those factors which are not significant at 90% confidence, the explanatory variables for this scenario were selected as: congestion on the target lane (Cgs t). Consequently, the utility function of changing lanes for this scenario is estimated as: DrvTypeB DrvTypeA Cgst LCV *781.0 *946.1*214.0314.3)( (6-10) Table 6-13 presents the number of maneuvers for different driver types that occurred during the tailgating scenario. Th e important factors iden tified for this scenario (see Table 4-9) are: factor 1: Travel speed, and the difference b/t travel speed and speed limit (Spd1, Spd1-Spd2); factor 2: Congestion on all lanes (Cgst1 and Cgst2); factor 3: The subject lane position (LanePos); and factor 4: Type of the lag vehicle (VehT). Table 6-13. Number of LCs for different driver types during Tailgating scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 9 6 5 1 Attempted 0 0 0 0 Completed 4 4 1 0 Total 13 10 6 1 In this case, only one tailgating maneuver was found for driver type D, which means this type of drivers doesnt get tailgating regula rly because of their higher aggressi veness. Or even if they get, they dont change lanes. Conseque ntly, the lane-changing probability for drivers with type D is

PAGE 152

152 set as constant 0. Only two dummy variables ( DrvTypeA and DrvTypeB ) were used in this model, and a utility function of changing lanes (d river types A, B and C) for this scenario is developed as follows: DrvTypeB DrvTypeA VehT Cgst Cgst SpdSpd Spd LCV *2*1*)21(*1* )(2 1 5 4 3 2 10 (6-11) With the field values obtained from 30 maneuvers (as shown in Table 6-13), the regression tool in SPSS was used to capture the relationship betw een these factors and th e probability value. The estimated results are pres ented in Table 6-14. Table 6-14. Estimated coefficients for the factors in Tailgating scenario Factors Coefficients T Value Constant -0.627 8.32 Subject travel speed (Spd1) -0.047 -0.66 Diff. b/t travel speed and spee d limit (Spd1-Spd2) -0.026 -2.32 Traffic congestion in the target lane (Cgst1) -0.018 -3.46 Traffic congestion in the subject lane (Cgst2) 0.029 -0.06 Subject lane position (LanePos) 0:left, 1: others -0.197 -12.39 Type of the lag vehicle (VehT) 0:car, 1: others 0.126 5.03 Driver Type A (DrvTypeA) 0.261 3.24 Driver Type B (DrvTypeB) 0.172 1.53 Among these parameters, the factor Subject lane position (LanePos) made the most significant contribution to the probability of ch anging lanes. The negativ e sign means drivers in left lane are more likely to change lanes to avoid tailgating (compared to the drivers in the median and right lanes), which is consistent with findings in the focus groups. Other significant factors include Diff. b/t travel speed and sp eed limit (Spd1-Spd2), Traffic congestion in the target lane (Cgst1) and Type of the lag vehi cle (VehT), which reflect the traffic dynamics on the road. As presented in Table 6-14, driver type B is not significant (at 90% confidence), and even though the driver type A is significant, the coefficient is only 0.261. By excluding the insignificant parameters, the explanatory variables fo r this scenario were selected as: Diff. b/t travel speed and speed limit (Spd1-Spd2), Traffic congestion in the target lane (Cgst),

PAGE 153

153 Subject lane position (LanePos) and Type of the lag vehicle (VehT). Consequently, the utility function of changing lanes for this scenario (driver types A, B and C) is estimated as: DrvTypeB DrvTypeA VehT LanePos Cgst SpdSpd LCV *172.0 *261.0 *126.0 *197.01*018.0)21(*026.0627.0)( (6-12) Table 6-15 presents the number of maneuvers for different driver types that occurred during the pavement scenario. The important factors identified for this scenario are: factor 1: Distance to the next turn (Dist); factor 2: Difference of the pavement (PavDiff); factor 3: Length of the pavement segment (LenPav); factor 4: Travel speed, and the difference b/t travel speed and speed limit (Spd1, Spd1-Spd2); and factor 5: Traffic congestion in the target lane (Cgst). Note: The factor 2 (PavDiff) is excluded from the estimation, since the value of factor is difficult to be collected from the in-vehicle test and is almost impossi ble to be modeled. Table 6-15. Number of LCs for different driver types during Pavement scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 4 12 7 5 Attempted 1 3 4 3 Completed 6 10 4 5 Total 11 25 15 13 Consequently, a utility function of changing lanes for this scenario is developed as follows: DrvTypeC DrvTypeB DrvTypeA Cgst SpdSpd Spd LenPave Dist LCV *)21(*1* **)(3 2 1 5 4 3 2 10 (6-13) With the field values obtained from 64 maneuvers (as shown in Table 6-15), the regression tool in SPSS was used to capture the relationship between these parameters and the probability value. The results are presented in Table 6-16. Among these parameters, the signif icant factors for this scenar io include Distance to the next turn (Dist), Length of the pavement segment (LenPav) and Traffic congestion in the target lane (Cgst), which refl ect the traffic dynamics on the road. The driver type seems to not

PAGE 154

154 affect the lane-changing decisions significantly in this scenario. As presented in Table 6-16, driver types A and B are not significant (at 90% confidence). Even though the driver type C is significant, the coefficient is not large (0.016). By excluding th e insignificant parameters, the explanatory variables for this sc enario were selected as: Dista nce to the next turn (Dist), Length of the pavement segment (LenPav) and T raffic congestion in the target lane (Cgst). Consequently, the utility function of changing lanes for this scenario is estimated as: DrvTypeC Cgst LenPav Dist LCV *016.0*153.0 *0015.0*005.0273.1)( (6-14) Table 6-16. Estimated coefficients fo r the factors in Pavement scenario Factors Coefficients T Value Constant 1.273 4.71 Distance to the next turn (Dist) -0.005 -1.74 Length of the pavement segment (LenPav) 0.0015 13.77 Subject travel speed (Spd1) -0.019 -0.32 Diff. b/t travel speed and speed limit (Spd1-Spd2) 0.112 0.63 Traffic congestion in the target lane (Cgst) -0.153 -4.92 Driver Type A (DrvTypeA) -0.548 -0.97 Driver Type B (DrvTypeB) -0.329 -1.32 Driver Type C (DrvTypeC) 0.016 2.14 By the end of the lane-changing probability model development, each preselected DLC scenario is related to a utility function of the respective factors and driv er types, as shown in Table 6-16. By reviewing the formula and coefficients used in these probability functions, with additional attention on the likeli hood results obtained from the focu s groups (see Table 4-8), it is believed that the trend of these functions ar e reasonable. By assuming the random components are independently and identically extreme value di stributed, the kernel of this binary choice model is logit, and the probability function of changing lanes for each DLC scenario can be calculated by Eq. (6-1). For the new added DLC scenarios (backup turning, pedestrian/scooter and erratic drivers), the number of maneuvers fo r different driver type s that occurred were presented in Tables 6-17 through 619. However, no utility function wa s obtained at this stage,

PAGE 155

155 Table 6-16. Scenario-bas ed utility functions ) ( LCV estimated from the in-vehicle data Scenario Utility functions and parameter R1:Stopped bus DrvTypeC DrvTypeB DrvTypeA Dist LocStop Cgst *139.1 *303.1 *533.2 *381.0 *116.19*236.048.6 R2:Vehicle merge DrvTypeC DrvTypeB DrvTypeA VehT Dist SpdSpd Cgst *166.3 *711.1 *0911. 1*09.1*018.0)21(*285.0*397.0398.4 R3:Slow vehicle ) (*382.0 *703.0 )21(*155.01*037.0*281.0*031.0743.3 DrvTypeC DrvTypeB DrvTypeA SpdSpd Spd Cgst Dist R4:Queue advantage DrvTypeC DrvTypeB DrvTypeA CurSig Dist QueDiff *277.0 *703.0 *694.1 *492.0*013.0 *254.0503.2 R5:Heavy vehicle DrvTypeB DrvTypeA Cgst *781.0 *946.1*214.0314.3 R6:Tailgating DrvTypeB DrvTypeA VehT LanePos Cgst SpdSpd *172.0 *261.0 *126.0 *197.01*018.0)21(*026.0627.0 R7:Pavement DrvTypeC Cgst LenPav Dist *016.0*153.0 *0015.0*005.0273.1 Table 6-18. Number of LCs for different driver types during Back Turning scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 5 7 3 2 Attempted 4 5 4 3 Completed 12 15 9 7 Total 21 27 16 12 Table 6-19. Number of LCs fo r different driver types during Pedestrian/Scooter scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 4 5 2 2 Attempted 3 1 4 3 Completed 7 10 8 5 Total 14 16 14 10 Table 6-20. Number of LCs fo r different driver types durin g Erratic Drivers scenario Maneuvers Type A(L1) Type B(L2) Type C(L3) Type D(L4) Potential 3 5 3 4 Attempted 4 6 7 3 Completed 10 13 15 9 Total 17 24 25 16

PAGE 156

156 since the significant factors for these DLC s cenarios were not generated in the focus group meetings. Consequently, the fixed probabilities ( ) obtained from the focus group study (see Table 4-8) were used for these scenarios. Further focus group studies may include these scenarios into the preselected DLC list, so that the important factors can be obtained. 6.2 Gap Acceptance Model for Urban Arterials In addition to the scenario-based probabil ity model, the gap acceptance procedure is studied in this section. A new algorithm was proposed to model the lane-changing gap acceptance into three modes (free, forced, and co operative/competitive). The free and forced lane changes are modeled as instantaneous events co nducted during the time interval immediately following the drivers decision, while the competitive/cooperative lane changes include more complex vehicle interactions. The emphasis is to model lane changes with interactions as a sequence of hand-shaking nego tiations between vehicles. Various interaction scenarios are modeled based on drivers actions and responses by referring to the nego tiation procedure used in computer network communi cations (Stevens, 1990, 1998). During the model formulation, first, differe nt lane-changing modes (free, forced, and cooperative/competitive) were identified from the field maneuvers, and the gap acceptance characteristics for different modes and driver types were assessed us ing field data. Next, notations used in the modeling framework are provided, followed by a set of quantitative criteria for distinguishing different t ypes of lane changes. Finally, scenarios related to the cooperative/competitive lane changes are analyzed in detail, and then are modeled with the corresponding values obtained from the in-vehicle data.

PAGE 157

1576.2.1 Gap Acceptance Characteristics This subsection presents the gap acceptance characteristics of the lane-changing related maneuvers obtained in the field. For the gap acceptance model development, only the completed and attempted lane changes are us ed (the potential lane changes are not considered). By studying a total of 601 completed lane changes and 199 attempted but unsuccessful lane changes, the maneuvers were classified into (i) free, (ii) forced, and (iii) competitive/cooperative based on previous studies (Hidas, 2005; Ben-Akiva et al., 2006; Choudhury, 2007; Sun et al., 2008) considering vehicle interactions as follows: Free lane change: there is no noticeable interaction be tween the subject and lag vehicle(s). The relative gap between the lead and lag vehicle(s) is large enough, so that the subject vehicle can move to the targ et lane with or without a change in its acceleration. Forced lane change: this type of lane change is followed by deceleration of the lag vehicle. Generally, the subject driver does not use turn signals or uses them very briefly before changing lanes. The lag vehicle does no t slow down until part of the subject vehicle has entered the target lane. Competitive/Cooperative (C/C) lane change: this type of lane change involves a sequence of interactions. First, the merging vehicle sends a lane-changing request to the lag vehicle by turning on the turn signal. The lag vehicle evaluates the request and may either cooperate by slowing down or not cooperate. The subject vehi cle re-evaluates the response based on the new gap and the speed of the lag vehicle. If the lane-changing criteria are satisfied, a cooperative lane change is executed. Otherwise, if the lane-changing request is rejected by the firs t lag vehicle, the subject vehicle would have to adjust its speed and re-send the request to the next follower. This process may last for several seconds, and the merging vehicle may give up the lane-changing attempt during the process. To distinguish between the forced and C/C lane changes, in addition to the turn signals, the spacing gaps between the subject vehicle and the lag vehicle we re recorded during a 6-second period, 3-second before and 3-second after the me rge. If the gap is incr easing before the entry point, it is a C/C merge. If the gap is either constant or narrow ing before the entry point, and starts to widen after merging, it is assumed that the subject vehicle has forced the lag vehicle to

PAGE 158

158 slow down. The maneuvers that occurred during th e in-vehicle experiment were grouped based on lane-changing modes for each given driver ty pe and are summarized in Table 6-21. In a total of 601 completed lane-changi ng maneuvers, 329 free, 124 forced and 148 competitive/cooperative lane changes (54.7%, 20.6 % and 24.6% respectively of the total lane changes) were observed. By relating the lane-c hanging modes to the driv er types, it was found that only two successful forced maneuvers we re conducted by type A drivers, which is reasonable since these drivers w ould be too timid to change la nes forcefully. All 199 attempted but unsuccessful lane changes belong to the co mpetitive/cooperative mode, since no rejection behaviors occurred during the free and forced maneuvers. As a results, the drivers of these maneuvers either changed lanes successfully (129 maneuvers for 64.8%), or gave up the attempt (70 maneuvers for 35.2%). Table 6-21. Number of co mpleted/attempted LCs based on modes and driver types LC Modes Type A Type B Type C Type D Total Free Completed 86 112 69 62 329 (54.7%) Forced Completed 2 22 58 42 124 (20.6%)Completed 31 50 36 31 148 (24.6%)Attempted 23 41 39 26 129 (64.8%)C/C Attempted 9 24 16 21 70 (35.2%) Completed 119 184 163 135 601 Total Attempted 32 65 55 47 199 Note: : The numbers indicate the unsuccessful lane -changing attempts included in situations which are finally successful. : The numbers indicate the unsuccessful lane-cha nging attempts included in situations which are finally aborted. Table 6-22 presents a summary of the acceptable spacing gaps measured from the completed lane changes, along with the rejected spacing gaps measured from the attempted but unsuccessful maneuvers. For the completed lane changes, the median gap length decreases from free to C/C, and then to forced lane changes fo r each of driver types, which means all drivers tend to accept smaller gaps from free to C/C, and to forced lane changes. However, by

PAGE 159

159 examining the spacing gaps across different driver types, no distinct tre nd was found. In addition to the spacing gap characteristics, vehicle acce lerations/decelerations were obtained from the completed lane changes as shown in Table 623. By examining the deceleration/acceleration values across different driver types, no significant differences were found. This means the acceleration/deceleration depends largely on th e drivers instantaneous perception on spacing/speed differences, which is not a good indi cator for reflecting the variation among driver types. With the numerical results for spacing gaps and accelerations presented, it is believed that although the type of subject driver somewhat affects the gap accep tance results, it doesnt affect the gap acceptance procedure as largely as in the lane-changing probability modeling. Consequently, the factor of driver type is onl y used in some special components, and doesnt affect the major decisions throughout the gap acceptance procedure. Table 6-22. Observed spacing ga p characteristics for the completed/attempted lane changes Gaps Type A (ft) Gaps Type B (ft) Gaps Type C (ft) Gaps Type D (ft) LC Modes max min med max min med max min med max min med Free Completed 5041 79 115 4691 83 105 4921 87 109 4771 85.5 111 Forced Completed 73.5 38.5 61 71 42.5 53 79.5 33.5 56.5 67.5 28.5 43 Completed 95 46 63 99.5 47 61.5 96 42 57.5 84.5 38 51 C./ C. Attempted 35.5 19 26 45 16 31.5 39.5 21 29 36 17.5 26.5 Note: : In the in-vehicle video, as some free lane changes dont ha ve lead/lag vehicle, such cases were not included in computing the statistics of the spacing gaps. Table 6-23. Observed accelerations/decel erations within vehicle interactions Subject vehicle, mph/s Lag vehicle, mph/s Driver Type # of Obs. Accel./ Decel. Max Min MedianMax Min Median Accel. 3.27 0.23 2.03 3.11 0.14 1.96 Type A 119 Decel. 4.71 0.17 2.61 4.93 0.16 2.71 Accel. 3.74 0.31 2.43 3.54 0.07 2.12 Type B 184 Decel. 5.23 0.06 2.72 5.19 0.04 3.06 Accel. 3.18 0.15 1.87 2.97 0.12 1.75 Type C 163 Decel. 5.12 0.28 2.44 5.41 0.13 2.93 Accel. 3.36 0.14 1.53 3.61 0.06 1.64 Type D 135 Decel. 4.73 0.17 2.07 5.31 0.12 2.87

PAGE 160

160 6.2.2 Notations and Modeling Framework Basic notations for describing vehicle inte ractions in the lane -changing process are illustrated in Figure 6-2. Vehicle S1 is the subject vehicle, which has an intention to merge into the target lane from the present lane. S0 is the vehicle in front of S1 in the present lane. T1, T2 and T3 are vehicles in the target lane, which may aff ect or be affected by th e lane-changing attempt of S1. T1 and T2 are the potential lead and lag vehicles, respectively. T3 is the lag vehicle following T2. If S1 is not able to merge in front of T2, then T1 or T3 becomes the potential lag vehicle. Gap1 is the initial existing spacing gap between T1 and T2, and Hlag is the spacing headway between S1 and T2. Figure 6-2. Initial scenario and notations adopted for the gap acceptance model Figure 6-3 presents the flowchart for the ga p acceptance m odel. Af ter accepting a lanechanging reason and deciding the targ et lane, the subject vehicle S1 checks the existing adjacent gap (gap1) to decide the lane-changing mode as: fo rced, free, C/C or ev en no lane-changing. Next, the lane-changing request is sent to the lag vehicle T2, only if the C/C mode is selected. Otherwise, a forced or free lane change is invoked, and the subject vehicle S1 will be moved to the target lane in the subs equent time interval. After T2 receives the lane -changing request S1, it

PAGE 161

161 has to decide whether to give wa y according to the spacing headway Hlag, along with the corresponding driver char acteristics. If courte sy is provided, a coope rative lane-changing is invoked, and S1 will re-check the adjacent gap (gap1) until a successful lane change becomes possible. Otherwise, competitive lane-changing is invoked. S1 will consider a forced lane change or adjust speed to consider a new merge (in front of T1 or T3). Under this framework, multiagent techniques were used to model the subject vehicle (S1) and the lag vehicles (T1, T2, T3, etc) on the target lane, and each vehicle is cons idered as an autonomous intelligent agent. The detailed communication scheme and reactive strategies from each of the vehicles are presented in sections 6.2.3 and 6.2.4. Competitive behaviorReject Accept Cooperative behavior S1re-checks existing gap1 for forced LCY N Forced lane-changing S1 checks gap1 for different LC modesC./C. T2evaluates the LC request based on Hlag S1moves to the target lane S1modifies speed, and resends LC request to T1or T3 S1gives up lane changing?N Y No lane-changing Free or Forced No lanechanging A LC reason is accepted & the target lane is decided Free or forced lane-changing Figure 6-3. Framework of the lane-changing algorithm

PAGE 162

1626.2.3 Decision Framework of the Subject Vehicle S1 As the first step of gap acceptance, rules are pr oposed in this section to distinguish between C/C lane changes and the other tw o types of lane changes, so that the free and forced lane changes can be modeled as an instantaneous even t, while the C/C lane changes are modeled as a negotiation process. The rules for distinguishing be tween C/C lane changes and th e other two types of lane changes are based on the initial gap on the target lane (gap1), critical gaps for different lanechanging modes, the su bject vehicle length (LS1) and the subject driver type. The output of this step is the probability of occurrence for each lane-changing mode. Figure 6-4 presents the situations in which different lane-changing modes may occur. There are six possible spacing intervals for gap1: I, II, III, IV, V and VI. For example, if gap1 falls into any intervals of I, II, IV, and VI, a deterministic maneuver occurs as no la ne-changing, forced la ne-changing, C/C lanechanging and free lane-changing, respectively. Otherwise, if the gap1 falls into the intervals III or V, two probabilistic alternatives may happen. Math ematical formulas and ru les for each situation, along with the output lane -changing mode(s), are expressed as follows. 0Max_Gapfree=+Min_GapfreeMin_GapC./C.Max_GapC./C.Max_GapforcedMin_Gapforced Forced lane changes C./C. lane changes Free lane changes I II III IV V VI Gap length Figure 6-4. Gap length on the target lane for different lane-changing modes Situation I: No lane-changing: Rule(s): gap1 < Min_gapForced, (6-15)

PAGE 163

163 Output: P(forced) = 0; P(C/C) = 0; P(free) = 0; Situation II: Forced lane-changing: Rule(s): gap1 Min_gapForced, (6-16) gap1 < min(Min_gapC/C, Max_gapForced), (6-17) Driver type A (6-18) Note: From the field driving te st, it was found that type A driv ers seldom conducted forced lane changes. Consequently, for this situation, if the dr iver type is A, the output will be the same as Situation I (no lane-changing) Output: P(forced) = 1; P(C/C) = 0; P(free) = 0; Situation III: Forced or C/C lane-changing: Rule(s): gap1 Min_gapC/C, (6-19) gap1 < Max_gapForced, (6-20) Output: C/C Forced C/C 1Min_gap Max_gap Min_gap )(P gap forced; (6-21) C/C Forced C/C 1Min_gap Max_gap Min_gap 1)/(P gap CC; (6-22) P(free) = 0; Situation IV: C/C lane-changing: Rule(s): gap1 max(Min_gapC/C, Max_gapForced), (6-23) gap1 < min(Min_gapFree, Max_gapC/C), (6-24) Output: P(forced) = 0; P(C/C) = 1; P(free) = 0; Situation V: C/C or Free lane-changing: Rule(s): gap1 Min_gapFree, (6-25) gap1 < Max_gapC/C, (6-26) Output: P(forced) = 0; free C/C free 1Min_gap Max_gap Min_gap )/(P gap CC; (6-27) free C/C free 1Min_gap Max_gap Min_gap 1)(P gap free; (6-28) Situation VI: Free lane-changing:

PAGE 164

164 Rule(s): gap1 max(Min_gapFree, Max_gapC/C), (6-29) Output: P(forced) = 0; P(C/C) = 0; P(free) = 1; where, gap1 is the initial gap on the target lane (as shown in Figure 6-1), Min_gapForced is the minimum distance for the forced lane changes, Min_gapC/C is the minimum distance for the cooperative and competitive lane changes, Max_gapForced is the maximum distance for the forced lane changes, Min_gapFree is the minimum distance for the free lane changes, Max_gapC/C is the maximum distance for the cooperative and competitive lane changes, Max_gapFree is the maximum distance for the free lane changes, assumed as + in this experiment, and LS1 is the vehicle length for the subject vehicle S1. After deciding the lane-changing mode, each type of maneuvers is modeled separately. If gap1 is too small to allow a lane change (
PAGE 165

165dT2 is the distance travel ed at the deceleration (DT2, driver type related parameter) that T2 would like to provide in this maneuve r, calculated from the formula: uT2*LCt0.5*T2D*2 LCt(uT2 is the speed for vehicle T2, LCtis the time for moving a vehicle from one lane to the adjacent lane), dS1 is the distance traveled by S1 during the lane change, calculated according to the formula: uS1*LCt (uS1 is the speed for vehicle S1), and gmin is the minimum safe constant gap between the subject vehicle and the lag vehicle, which is independent of the speed difference. Figure 6-5. Effective components in cluded in the lag spacing headway Formula 6-30 indicate s that the existing Hlag (before merge) ensures the spacing headway between S1 and T2 larger than min S1Lg at the end of the merging maneuver. Here it assumes the speed of S1 is constant, and T2 uses T2D deceleration. Thus, the decision tree for the process of competitive/cooperative behavior is illustrated in Figure 6-6, followed by descriptions of the two strategies.

PAGE 166

166 Cooperative behavior T2 maintains speed T2 accelerates T2 decelerates Lane-changing request accepted by T2 Lane-changing request rejected by T2 S1 checks for T1or T3 S1 performs forced lane change S1 performs coop. lane change Competitive behavior Forced(DT) >Rand() NT2(DT)= C or D Hlag, driver types (DT) for T2and S1T2(DT)= A or BY Figure 6-6. Decision tree for modeling competitive/cooperative behavior 6.2.4.1 Competitive behavior As shown in the left part of Figure 6-6, T2 chooses to reject the lane-changing request as Hlag is unacceptable. In this case, the response of T2, as presented in Figure 6-7, needs to be reevaluated. Two possible scenarios may occur: 1) T2 maintains its speed or 2) T2 decides to accelerate. Detailed analyses on th e two potential sequences of T2 are given as follows. Present lane Target lane S1 T1 T2 T3 gap2Hlag (a) (b) Figure 6-7. Competition scenario in th e competitive/cooperative lane changes (a) Negotiation with the immediate lag vehicle (b) Lane-changing communication scheme T2 maintains speed: A normal/defensive driver of T2 (driver types A or B) would reject the lane-changing request, and maintain the curren t car-following state. In this case, S1 either

PAGE 167

167 attempts a forced lane change, or considers T1/T3 as the new candidate lag vehicle and adjusts speed accordingly. The driver type of S1 was used to decide the probability for attempting a forced lane change. The conducting probabilities of the four driv er types were calculated from the field data (in Table 6-21) as follows: Forced(A) = 2/33 = 6.1% Forced(B) = 22/72 = 30.1% Forced(C) = 58/94 = 61.7% Forced(D) = 43/72 = 59.7% In this calculation, only the number of forced lane changes and C/C were considered, since the traffic situation is far from the definition of free lane changes. Th en, if a forced lane change is initiated, the new speed for S1 is calculated using the car-f ollowing formula by adopting the T1 as the lead vehicle. Otherwise, S1 considers T1 or T3 as the new candidate lag vehicle and adjusts speed accordingly. If S1 considers T3 as the candidate lag vehi cle, the new speed for S1 is calculated as: T2 S1 max S1 old 2Tmax old S1 T2 S1 max S1 old 2Tmax old S1 new S1 )*b min( )*b max( uu D uAu uu D uDu u (6-31) where new S1u is the new speed for the vehicle S1, old S1u is the initial speed for the vehicle S1, old T2u is the initial speed for the vehicle T2, maxD is the maximum deceleration for the given traffic flow, Amax is the maximum acceleration for the given traffic flow, and bS1 is the driver aggressiveness related deceleration parameter for S1, and will be calibrated in the model implementation. Eq (6-31) indicates S1 tries to decelerate to a speed value at max S1*b D lower than T2, so that it can attempt a lane changing in front of T3. If S1 considers T1 as the candidate lag vehicle, the new speed for S1 is calculated as:

PAGE 168

168 T2 S1 max S1 old 1Tmax old S1 T2 S1 max S1 old 1Tmax old S1 new S1 )*b min( )*b max( uu D uAu uu D uDu u (6-32) where old T1u is the initial speed for the vehicle T1, all other notations are as Eq. (6-31). Eq (6-32) indicates S1 tries to accelerate to a speed value at max S1*b D higher than T1, so that it can attempt a lane changing in front of T1. T2 accelerates: An aggressive driver of T2 (driver types C or D) w ould reject the lane-changing request and accelerate so th at the subject vehicle S1 cannot force its way into the target lane. In this situation, T1 and T3 would keep the previous car-following state, while T2 accelerates so that the new gap (gap2 in Figure 6-7a) reaches the value b1*(gmin+LS1) (b is the driver aggressiveness related gap acceptance parameter) to impede the merging maneuver. The new speed for T2, new T2u is then calculated as: LC St Lgbgap A uu )(* ,min1 min 12 max old T2 new T2, (6-33) where new T2u is the new speed for the vehicle T2, b1 is the driver aggressiveness related gap accepta nce parameter, and will be calibrated in the model implementation. In this case, S1 has to consider T3 or T1 as the new candidate lag vehicle and adjust speed by using Eq.s (6-31) and (6-32) accordingly. 6.2.4.2 Cooperative behavior Instead of competing with the merging vehicle, as shown in the right part of Figure 6-6, T2 may accept the lane-changing request, if Hlag is acceptable. In this case, T2 responds to the request by decelerating, as presented in Figure 6-8, and S1 tries to accelerat e to the center position of gap2. The new speed for T2 and S1 can be calculated as:

PAGE 169

169 S1 T2 max T2 old T2 S1 T2 max T2 old S1max old T2 new T2 *b )*b, max( uu D u uu D uDu u (6-34) and 0 ), max( 0 ), min(old S1max old S1 old S1max old S1 new S1d duDu d duAu u (6-35) where, new T2u is the new speed for the vehicle T2, old T2u is the initial speed for the vehicle T2, bT2 is the driver aggressiveness related deceleration parameter for T2, and will be calibrated in the model implementation, and d is the distance from the position of S1 to the center position of gap2, all other notations are as Eq.s (6-31) and (6-32) (a) (b) Figure 6-8. Cooperation scenario in the compe titive/cooperative lane changes (a) Negotiation with the immediate lag vehicle (b) Lane-changing communication scheme 6.3 Summary and Conclusions In this chapter, two key components of th e lane-changing model, the scenario-based probability model and the gap acceptance model, are presented. In the scenario-based probability model, driv ers actions under different DLC scenario are examined and modeled. The acceptance of each pa rticular lane-changing scenario (scenariobased probability) is modeled as a function of the corresponding important factors and driver types. The modeling parameters were estimated using a regression method (logistic regression) with detailed lane-changing data obtained in th e field. The driver type classification scheme

PAGE 170

170 found in the focus groups and confirmed in the invehicle experiment was incorporated, so that the lane-changing behaviors for different type of drivers can be modeled. The proposed algorithm enumerates all DLC scen arios occurring in urban arteri als, and it models each one probabilistically. In the gap acceptance model, the three lane-changing modes (free, forced and competitive/cooperative) are modeled with an em phasis on capturing vehicle interactions during lane changing. For the free and forced lane change s, the subject vehicle is moved to the target lane, and the car-following strategy is applied subsequently to the corresponding vehicles. The procedure of the competitive/cooperative lane changes is modeled as a sequence of handshaking negotiations with more complex interactions. This approach differs from existing models that assume that lane changes are alwa ys conducted instantaneousl y or within fixed time intervals and the different lane -changing modes are not intercha ngeable. In the proposed model, the games between the subject vehicle and the lag vehicle may be a co mpetition or cooperation depending on the characteristics of the surrounding traffic and drivers. Under this modeling framework, the strategies of not change, fr ee change, C/C change and forced change are interchangeable, which better reflects the reality of urban arterial lane changes. To implement the proposed lane-changing mode l and validate the capabilities of the new algorithm, the selected parameters for the gap acceptance model were estimated from the invehicle field data. The gaps and acceleration valu es, as presented in Table 6-17 and Table 6-18, along with the other information obt ained during the driving tests, were used to initialize the parameter settings for the lane-changing gap acceptance model as follows: Minimum distance for the forced lane changes, Min_gapForced: 36 ft, Maximum distance for the forced lane changes, Max_gapForced: 73 ft, Minimum distance for the C/C lane changes, Min_gapC/C: 44 ft,

PAGE 171

171 Maximum distance for the C/C lane changes, Max_gapC/C: 94 ft, Minimum distance for the free lane changes, Min_gapFree: 83 ft, Maximum distance for the free lane changes, Max_gapFree: assumed as + Vehicle length for the subject vehicle S1 (Honda Pilot), LS1: 16 ft, Time for moving vehicle from the pr esent lane to the adjacent lane, LCt : 1.0 sec, Minimum safe constant gap, gmin: 18 ft, Maximum deceleration, Dmax: 5 mph/s, Maximum acceleration, Amax: 3 mph/s, Driver aggressiveness rela ted deceleration parameters, S1b or T2b: 2.8/5 = 0.56, and will be calibrated in the model implementation Driver aggressiveness related gap acceptance parameter, b1: 1.0, and will be calibrated in the model implementation. The unobserved parameters, such as driver aggressiveness related parameters (S1b T2band 1b), were initialized based on the relationship inferred from field data.

PAGE 172

172 CHAPTER 7 MODEL IMPLEMENTATION AND VALIDATION This chapter presents the im plementation a nd validation details of the developed lanechanging model, as presented in Chapter 3 (Figur e 3-5). Section 7.1 presents the validation field dataset collection and analysis effort. Next, in Section 7.2, the proposed model is implemented as an RTE (run time extension) plug-in in CORSIM followed by an aggregate calibration to tune up the selected behavioral parameters within the simulation model. Section 7.3 presents the systematic validation to evalua te the agreement between the simulation results (for both new and CORSIM original lane-changing models) and the field observations based on the selected indices of measures. The chapte r concludes with the results fro m various analyses designed to test the effectiveness of the newly developed lane-changing model and a summary of the findings. 7.1 Datasets As the first step of model validation, a heavil y congested arterial segment in Gainesville, FL was selected. The segment is on Newberry Road, stretching from I-75 on the west to the main Oaks Mall entrance toward the east, with a di stance of about 1,650 ft (Figure 7-1). The posted speed limit for the three-lane segment is 35 mph. Characteristics of the arterial are the high daily traffic attracted by the adjacent shopping center (O aks Mall) and closely spaced intersections in the vicinity, which result in a high number of lane changes. Drivers are free to select the lane with the highest utility as the target lane and make subsequent lane changes depending on availability of gaps along th e stretch of the arterial.

PAGE 173

173 Figure 7-1. The Newberry Road segment for data collection (source : Microsoft Bing Maps) A sketch of the data collection site with th e lane channelization is shown in Figure 7-2. Video data were collected from the arterial segment by using traffic surveillance cameras mounted on street lamp poles along the roadway. I-75 C C C C Oaks Mall West NW 69th Terrace NW 66th Street NB Ramps EB WBC : location for mounting the surveillance cameras Newberry Rd. Figure 7-2. Sketch of the segment selected for validation data collection (not to scale) The Newberry Road video data were collected along the arterial on May 3rd, 2005, from 10 am to 6 pm (Washburn and Kondyli, 2006). From approximately 32 hours of video recording, eight hours of data during heavy traffic conditions are selected to study lane-changing behaviors.

PAGE 174

174 A total of 138 successful lane-changing maneuvers are identified in the vi deo, including 79 free, 34 forced and 25 C/C lane changes (57.2%, 24. 6% and 18.1% to the total lane changes respectively). Table 7-1 presents the summary of the acceptable spac ing gaps and vehicle acceleration/deceleration that occurr ed within the three types of lane changes. Similar to the invehicle gaps, the median gap lengt h decreases from free to C/C, and then to forced lane changes, while no significant difference was found in accelera tion/deceleration. This further confirms that the acceleration/deceleration depends largely on the drivers instantaneous perception on spacing/speed differences. However, the field co llected acceleration/dece leration values provide a quantitative range for simulation parameter estima tion. Only the traffic flow related datasets were obtained, since characteristic s of the drivers, such as aggressiveness, were not available. Table 7-1. Summary of the Newberry Road video data (a) Observed spacing gaps on the adjacent lane (b) Observed acceleration/dece leration for the subject and lag vehicles (a) Adjacent Spacing (ft) LC Type # of Obs. Max Min Median Free 79 324 89.5 133 Forced 34 91.5 37 54.5 C/C 25 102 42.5 72 (b) Acc. (subject veh, mph/s) Acc. (lag veh, mph/s) LC Type # of Obs. Max Min Median Max Min Median Free 79 2.5 -1.3 0.2 1.9 -0.8 0.1 Forced 34 4.5 -3.0 -0.3 -2.2 -6.9 -4.6 C/C 25 3 -5 -0.6 2.8 -5 -0.4 Note: : In the video, as some lane changes dont have lead/lag vehicle, such cases were not included in computing the statistics of the spacing gaps. 7.2 Model Implementation and Calibration This section presents the lane-changing m odel implementation and validation efforts, referred as stage I (Model Implementation and Ca libration) in Figure 3-5. First, the Newberry Road network was simulated based on the field da ta. Calibrations were conducted to make sure

PAGE 175

175 the simulated network replicat ed the actual traffic conditions Next, the new lane-changing model was implemented as an RTE plug-in, whic h is invoked dynamically during the simulation to replace CORSIMs default lane-changing strate gies. Finally, an aggregate calibration is conducted to tune up the selected behavioral pa rameters within the newly developed model. 7.2.1 Model Implementation The traffic volume data obtained from the vi deo were used for simulating the network in CORSIM, and these pertain to mainline and crossstreet volumes as well as percentages of heavy vehicles. Figure 7-3 contains hourly volume da ta along the study segment, which were obtained through the data reduction. As described by Wa shburn and Kondyli (2006), four categories of vehicles were recorded: passenger car, medium truck, large truck and bus. The truck and bus categories were combined to obtain a heavy vehicle percentage for each approach along the arterial. Signal timing data for each intersection were obtained from the City of Gainesville, which were further confirmed from the video observation. Travel times for both directions between I-75 NB ramps and NW 66th St. were obtained throu gh vehicle matching (manual observation of vehicles in video), and were used for model calibration. Figure 7-3. Volume data from video reduction taken on May 3rd 2005 PM peak period (Washburn and Kondyli, 2006)

PAGE 176

176 As the first step of model calibration, th e Newberry Road network was simulated and calibrated using CORSIMs default lane-changing al gorithm (for 15 runs). Th e calibration at this stage was performed by adjusting simula tion parameters settings such that: 1) The CORSIM simulated travel times were within +/10% of the field-measured travel time for both approaches (WB and EB), and 2) The total numbers of lane-chang ing maneuvers were within +/ 20% of the field-measured values for both approaches (WB and EB). Table 7-2 presents CORSIMs lane-changing parameters and LC-related driver behavior parameters chosen for this calibration, along with their initial a nd calibrated values. The values of the parameters were adjusted mainly base d on the field collected lane-changing maneuvers. Table 7-3 presents the average travel time and number of lane changes for both approaches (WB and EB) by 15 simulation runs, as well as the corresponding field data values. As seen, the simulated travel times in both approaches were successfully adjusted into +/-10% of the field measured values, and the numbers of lane-c hanging were within +/-20%, which mean all calibration criteria were met. Table 7-2. Initial and calibra ted values of the parameters in CORSIM lane-changing model Parameter Value Calibrated Parameter Initial Calibrated Duration of LC maneuver 3 sec 2 sec Min deceleration for a LC 5 ft/sec*sec 7.5 ft/sec*sec MLC 10 ft/sec*sec 9 ft/sec*sec Diff. in min/max acceptable dece. DLC 5 ft/sec*sec 6 ft/sec*sec Headway all drivers will attempt a LC 2.0 sec 1.5 sec Headway no drivers will attempt a LC 5.0 sec 3.0 sec Table 7-3. Travel time and numbe r of lane-changing measurements Measurements Simulated Fieldmeasured Error WB 53.7 sec 59 sec -8.98% Travel Time EB 73.4 sec 67 sec 9.55% WB 26 32 18.7% Number of LCs EB 37 44 15.9%

PAGE 177

177 In the next step, the new lane-changing mode l was implemented as a C++ plug-in (.dll), which interfaces with CORSIM engine during th e simulation. Commands in TShell environment were used for the RTE deployment (FHWA, 2006). In CORSIM v6.1, as shown in Figure 7-4, for each simulation time step, CORSIM Server calls a series of functions within CORSIM to drive the simulation even t loop (FHWA, 2006). The RT_PRE_NETSIM_VEHICLE message is sent just prior to calling the FORTRAN subr outine MOVE, which handles lane-changing, carfollowing, etc., to move all the vehicles for th e current time step. The lane-changing plug-in is set up to respond to that message, and the func tion within the plug-in would perform the lanechanging maneuver. With the completion of lane -changing function, the CORISM lane-changing timer is set to a value that would prevent the embedded lane-changing logic from being applied. The subroutine MOVE would still be called, but vehicles would not be allowed to make a lane change. A configuration file (.cfg) is used to st ore the values of the calib rated driver behavioral parameters for the model implementation. Figure 7-4. CORSIM entire architecture and the communication with lane-changing plug-in (source: FHWA, 2006; modified by the author)

PAGE 178

178 Figure 7-5 presents the flow of the lane-cha nging decision implementation procedure. The main function within the plug-in first tries to determine whether any DLCs or MLCs may be invoked. The lane-changing probability for a ML C is always 1. For a DLC, the corresponding probability function is used to calculate the probability of changing lanes. Figure 7-5. Implementation of the lane -changing decision procedure in CORSIM

PAGE 179

179 When a lane change is decided, the rules (6 -15) (6-29) are used to distinguish the maneuver as free, forced or C/C or no lane change. The free and forced lane changes are modeled as an instantaneous event by simply movi ng the subject vehicle to the target lane during the following time interval (t+1), so that CO RSIMs default car-following algorithm can be applied to the subject vehicle on the target lane For the C/C lane changes, the strategies as proposed in Eq.s (6-30) through (6-34) are used to model the maneuver as a sequence of handshaking negotiations. Afterwards, by setting CORSIMs lane-changing timer to a value that would prevent the embedded lane-changing logic from being applied, the subroutine MOVE would still be called, but no vehicl e is allowed to change lanes. During this implementation, only the pre-selected lane changes were captured and modeled by the RTE plug-in. Other types of lane changes are ignored, and CORSIM embe dded lane-changing logic is invoked to handle those issues. Practically, CORSIM sets the maximum number of lanes for any link as 7, and a global index for any given lane K on road link IL is ca lculated as: ILL = (IL-1)*7+K. All vehicles in this lane (K) are stored in a double-linked list da ta structure, which can be exported and accessed by the RTE plug-in. As illustrated in Figure 7-6, in modeling a lane change, the subject vehicle has to be removed from the original lane linke d list and inserted into the corresponding position on the target lane link list. Both operations are co mpleted in the same simulation time step (t+1). First, the subject vehicle S1 is deleted from the original list by connecting S2 directly to S0. Next, the connection between T1 and T2 is broken, and both vehi cles are connected to S1. Once the new connections between vehicles are set, the accelera tions of all vehicles ar e calculated by the carfollowing model in CORSIM, with respect to the lead vehicle in the same lane.

PAGE 180

180 Figure 7-6. Schematic representati on of a lane change in CORSIM Another important issue during this implementation is how to assign CORSIMs 10 driver types to the proposed four groups. As discussed in Chapter 5, the behavior-based index, FAI, is used to categorize the 40 driving participants into four groups as A (<= 4.2), B (4.3 5.7), C (5.8 6.8) and D (>= 6.9). In CORSIM simulation, 10 different driver t ypes, ranging from most passive (i.e., driver type 1) to the most aggressive (i.e., driv er type 10), are gene rated to represent driving behavioral characteristic s. Driver type for each CORSIM vehicle is randomly selected from the embedded discrete uniform distributi on (Chien, et al., 2001), which means each type accounts 10% of the total number of driver. Initially the above FAI values we re used directly to classify all CORSIM simulation drivers in to four groups, as shown in Figure 7-7. Figure 7-7. CORSIM driv er classification based di rectly on the FAI values

PAGE 181

181 According to the three boundary points (4.2, 5.7 and 6.8), CORSIM simulation drivers can be easily classified into four groups. The percentage of drivers for each corresponding type (A, B, C and D) in the simulation was calculated as 42%, 15%, 11% and 32%, respectively. This contradicts the actual percentages of driver types, wherein the percentages of drivers for A, B, C, and D are 9/40 = 22.5%, 12/40 = 30%, 11/40 = 27.5%, and 8/40 = 20%. Consequently, the percentages from the sampli ng participants for each driver type were used to assign CORSIMs 10 driver types to the dr iver types A, B, C, and D used in this study. As shown in Figure 7-8, the type A has 22.5% of all drivers, and the types B, C, and D have 30%, 27.5%, and 20%, respectively. Three bound ary points are set as 2.25, 5.25 and 8, respectively. Figure 7-8. CORSIM driver classification based directly on the sampling percentages Table 7-4 presents the final assignment of driver types. All CORSIM types 1 and 2 drivers were assigned to type A. CORSIM type 3 drivers were split into two parts, 25% was assigned to type A, and the rest 75% goes to type B. The types 4 and 5 drivers were assigned to type B. 25% of the type 6 drivers were assigned to type B, and th e rest 75% were assigned to type C. All types 7 and 8 drivers were assigned to type C and types 9 and 10 drivers we re assigned type D. Table 7-4. Strategy for di stributing CORSIM drivers in to the lane-changing groups CSM drv Clusters 1 2 3 4 5 6 7 8 9 10 Type A (%) 100 100 25 0 0 0 0 0 0 0 Type B (%) 0 0 75 100 100 25 0 0 0 0 Type C (%) 0 0 0 0 0 75 100 100 0 0

PAGE 182

182Type D (%) 0 0 0 0 0 0 0 0 100 100 7.2.2 Aggregate Calibration To be distinguished from CORSIM origina l model, the newly developed lane-changing model is referred to as the new model. In this stage, the estimated coefficients and parameters for the new model were included as the init ial settings of the CORSIM plug-in. Given that CORSIMs lane-changing and behavior-related parameters have already been calibrated, the following behavior-related parameters in the new model were selected for calibration: Minimum safe constant gap, gmin, Maximum deceleration for the given traffic, Dmax, Maximum acceleration fo r the given traffic, Amax, Driver aggressiveness related deceleration pa rameter for the subject vehicle in the lane change, S1b Driver aggressiveness related deceleration para meter for the lag vehicle in the lane change, T2b, and Driver aggressiveness related gap acceptance parameter, b1. During this calibration effort, each gap or acc eleration/deceleration parameter was set to change from -50% to +50% while keeping the re st of the parameters fixed (Ben-Akiva et al., 2004). Each driver aggressive ness related parameter (S1b ,T2band b1) was changed from 0.2 to 1.2. The group of parameters that generates travel tim es and number of lane changes closest to the field-measured values was identified for the si mulation and validation followed. By the end of the adjustment, the average simulated travel ti mes for the new model were obtained as 56.1 sec and 69.8 sec for westbound (WB) and eastbound (EB) tr affic, respectively, which are within the +/5% range of the field-measured travel time. The numbers of lane-changing are also within +/20% for both approaches (37 for WB, and 50 for EB). As shown in Table 7-4, most of the parameters do not change significantly during calibration except the minimum safety gap for lane changes (gmin), and the driver aggressiveness related gap acceptance parameter ( b1). This is

PAGE 183

183 expected since the calibration datasets (collect ed from Newberry Road segment) and the invehicle datasets (collected from the Newberry route and the Waldo route) are similar to each other both in the time-of-day and road geometry, as well as driver characteristics. The only two big differences are: 1) the posted speed limit for Newberry Road is 35 mph, but the speed limits for the in-vehicle route change from 30 mph to 45 mph in its different segments; and 2) for the Newberry Road WB traffic, the number of left turns is much higher than right turns during the peak hours because of the large attraction of the Oaks Mall Shopping Center. The calibration parameters and their before-and-aft er values are listed in Table 7-5. Table 7-5. Initial and calibra ted values of the parameters in the new lane-changing model Parameter Value Calibrated Parameter Index Initial Calibrated Minimum safe gap gmin 18 ft 14 ft Maximum deceleration Dmax 5 mph/s 5.5 mph/s Maximum acceleration Amax 3 mph/s 3 mph/s Decel. parameter for S1 bS1 0.56 0.60 Decel. parameter for T2 bT2 0.56 0.50 Aggr. related parameter b1 1.0 0.9 7.3 Model Validation The purpose of system validation is to test a nd determine the extent to which the simulation model replicates the real syst em under different traffic cond itions (Toledo and Koutsopoulos, 2004; Ramanujam, 2007; Ramanujam et al., 200 8). The following section presents the comparison between the simulated results of the two models (CORSIM an d new) and the field data, referred as stage II (Model Validation) in Figure 3-5. Both calibrated models were simulated with OD demands measured from different day (April 30th, 2005) video data under congested traffic conditions, as presented in Figure 7-9. Three measures of performance, average lane-based travel time, vehicles lane distributi on, and cumulative lane changes by vehicles, were selected to evaluate the model performance because of their close relevance to the lane-changing

PAGE 184

184 behavior. Finally, a sensitivity analysis (Badra, 2007) was conducted to demonstrate how the variation in the output of the simulated model can be apporti oned quantitatively. Various goodness-of-fit measures were used to evaluate the overall performance of both simulation models. Figure 7-9. Volume data from video reduction taken on April 30th, 2005 PM peak period (Washburn and Kondyli, 2006) 7.3.1 Comparison of the Lane-Based Travel Time As introduced previously, the Newberry Road se gment is divided into three sections by the four signalized intersections. The lane-based average travel time was obtained from the video data by matching vehicles at the entrance and exit of each arterial secti on manually. Traffic flows on approaches for both directions were used to obtain the average travel time. Figure 7-10 shows the comparison of average la ne-based travel time between the new model, the validation field data, and the origi nal model for each section. From the results shown, we found that the original model tends to underestimate the travel time. By incorporating the new model, travel time, especially for the right and the left lanes, becomes closer to the field data. Additionally, the diffe rences in the by-lane travel time are more significant after applying the new strategy, and closer to the fi eld data than the original CORSIM simulation. Although the average travel times are similar, the new model gives a better representation of the lane-by-lane differences.

PAGE 185

185 Comparison of Travel Time (Section 1 WB)0 2 4 6 8 10 12 14 16 18 1234Average TT (sec.) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) Average Comparison of Travel Time (Section 1 EB)0 5 10 15 20 25 30 1234Average TT (sec.) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) Average (a) (b) Comparison of Travel Time (Section 2 WB)0 5 10 15 20 25 30 35 40 45 50 1234Average TT (sec.) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) Average Comparison of Travel Time (Section 2 EB)0 5 10 15 20 25 30 35 40 1234Average TT (sec.) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) Average (c) (d) Comparison of Travel Time (Section 3 WB)0 5 10 15 20 25 30 1234Average TT (sec.) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) Average Comparison of Travel Time (Section 3 EB)0 2 4 6 8 10 12 14 16 1234Average TT (sec.) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) Average (e) (f) Figure 7-10. Comparison of the lane-based av erage travel time (a) Section 1 westbound (b) Section 1 eastbound (c) Section 2 westbound (d) Section 2 eastbound (e) Section 3 westbound (f) Section 3 eastbound

PAGE 186

186 The two algorithms (new and original) were compared in relation to the validation field data. Chi-Square tests were performed between the expected travel time (f rom field data) and the two observed travel times (from simulation) for both WB and EB traffic, as shown in Table 7-6. For the WB traffic, the simulated travel time s for both cases (new and original) were not statistically different from th e field travel times, with the 2 value as 2.107 and 4.370 respectively. However, there is greater level of confidence associated in this statement when using the new algorithm (97% for 2=2.107) than using the existing CORSIM algorithm (82% for 2=4.370). Similar Chi-Square results were obt ained from the EB traffic as that the simulated travel times for both new and original cases were not statistically different from the field travel times, with the 2 value as 2.695 and 5.955 respectively. Result from the new algorithm is with higher level of confidence 95% for 2=2.695 instead of 65% for 2=5.955. Table 7-6. Comparison of simulation travel time between new and CORSIM models (2 Test) T.T. (sec.) WB T.T. (sec.) EB ( Oi Ei) /Ei WB ( Oi Ei) /Ei EB Location New CORSIM Field NewCORSIMField New CORSIM New CORSIM Section 1 Lane1 15.3 12.7 14 24.4 21.8 26 0.121 0.121 0.098 0.678 Lane2 11.7 12.3 11 18.4 19.5 19 0.045 0.154 0.019 0.013 Lane3 17.5 14.1 16 24.3 19.2 21 0.141 0.226 0.519 0.154 Section 2 Lane1 44.8 39 42 30 25 34 0.187 0.214 0.471 2.382 Lane2 39.7 39.4 37 20 16.6 17 0.197 0.156 0.529 0.009 Lane3 45.9 37.1 44 24 28 26 0.082 1.082 0.154 0.154 Section 3 Lane1 23.4 21.6 26 10.5 12.5 8 0.260 0.745 0.781 2.531 Lane2 19.9 17.3 18 13.9 13.4 13 0.201 0.027 0.062 0.012 Lane3 17.5 18.8 14 8.7 8.4 8 0.875 1.646 0.061 0.020 Overall Value (2) 2.107 (97%) 4.370 (82%) 2.695 (95%) 5.955 (65%) Two-sided T tests were performe d to investigate whether (i) la ne-based travel time of new model is equal to that of the fi eld data, and (ii) lane-based trav el time of original model is equal to that of the field data. Results of the firs t test showed no evidence that the travel time is

PAGE 187

187 different for all travel time at a 95% confidence level except the EB right lane travel time in section 3. The second test showed that the EB ri ght lane travel time (for Section 3) and the WB left lane travel time (for all three sections) ar e significantly different from the field-measured values at a 95% confidence level. In addition, a comparison between the two al gorithms was performed as shown in Table 77. An F test was first conducted to compare variances, and it was found that in all but one case (4.372 > F(95%) = 2.480, Section 3 WB, Lane 1 at Table 7-6) the variances are the same. Since the sample size was relatively small (15 simulation runs), T test with equal and unequal (and unknown) variances was selected to compare the means at a 95% confidence level. Table 7-7. Comparison of simu lation travel time between new and CORSIM models (T Test) Mean T.T. (sec.) St. Dev. (sec.) Location New CORSIM New CORSIM Pooled F Test T Test Is T.T. statistically different? Section 1 WB Lane1 15.3 12.7 2.66 2.93 2.798 1.213 2.545 YES Lane2 11.7 12.3 2.71 2.63 2.670 1.062 0.615 NO Lane3 17.5 14.1 3.12 2.05 2.640 2.316 3.527 YES Section 1 EB Lane1 24.4 21.8 3.13 3.72 3.438 1.413 2.071 YES Lane2 18.4 19.5 2.67 3.26 2.980 1.491 1.011 NO Lane3 24.3 19.2 3.32 2.49 2.934 1.778 4.760 YES Section 2 WB Lane1 44.8 39 5.07 5.96 5.533 1.382 2.871 YES Lane2 39.7 39.4 5.34 5.62 5.482 1.108 0.150 NO Lane3 45.9 37.1 7.23 6.94 7.086 1.085 3.401 YES Section 2 EB Lane1 30 25 4.77 4.06 4.429 1.380 3.092 YES Lane2 20 16.6 4.23 5.01 4.636 1.403 2.008 YES Lane3 24 28 3.61 4.49 4.074 1.547 2.689 YES Section 3 WB Lane1 23.4 21.6 4.83 2.31 3.786 4.372 1.302 NO Lane2 19.9 17.3 4.32 5.34 4.857 1.528 1.466 NO Lane3 17.5 18.8 2.06 1.49 1.798 1.911 1.980 YES Section 3 EB Lane1 10.5 12.5 1.83 2.31 2.084 1.593 2.628 YES Lane2 13.9 13.4 2.32 2.34 2.330 1.017 0.588 NO Lane3 8.7 8.4 1.56 1.79 1.679 1.317 0.489 NO The comparison results show that by introducing the new model, the changes of the travel time on the left lane and right lane ar e different for all three sections (except Lane 1 in the Section 3

PAGE 188

188 WB, and Lane 3 in Section 3 EB). However, the travel time for the middle lane (except Lane 2 in the Section 2 EB) does not differ much between the two algorithms. One potential explanation is that most defensive or unwilling to change lane drivers would keep driving in the middle lane, which induces that the new lane -changing model doesnt affect vehicles on the middle lane as much as those on the left/right lanes. 7.3.2 Comparison of the Lane Distribution The vehicle lane distributions were obtained from the video data a nd compared with the simulated counterparts. In bot h simulations implemented in CORSIM, surveillance detectors were placed on each lane for every 50 feet to r ecord the number of vehicles using the lane. The field lane utilizations were observed from a ll four locations, and aggregated to obtain the percentages of the traffic distributions on each lane. Figure 7-11 shows a comparison of lane distribution from the new model, the valida tion field data and the original model. Figure 7-11a presents the simulation results of the new model and the original model for the WB traffic, which are similar to each ot her. Both simulations tend to overestimate the utilization of the middle lane (lane 2), and underes timate the utilization of the left lane (lane 3). However, the results from the new model are cl oser to field observation s for both lanes (lanes 2 and 3). The root mean square error (RMSE), as de fined in Eq. 7-1, is calculated for the vehicle lane distributions in CORSIM original m odel and the new lane model as 0.051 and 0.0374 respectively, which indicate an improvement of 26.62 %. 2 3 1)( 3 1 i obs i sim iYY RMSE (7-1) where: i is the lane index, (1: right lane, 2: middle lane, a nd 3: left lane),

PAGE 189

189sim iY and obs iY are the observed and simulated percentage of vehicle lane dist ributions on lane i. Comparison of Lane Distribution (for Entire Arterial WB)0 5 10 15 20 25 30 35 40 123Percentage (%) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) (a) Comparison of Lane Distribution (for Entire Arterial EB)0 5 10 15 20 25 30 35 40 45 123Percentage (%) New Field Data CORSIM Lane 1 (Right) Lane 2 (Middle) Lane 3 (Left) (b) Figure 7-11. Comparison of the la ne distribution (a) Lane distri butions for the WB traffic (b) Lane distributions for the EB traffic Figure 7-11b presented the results of the new model and the origina l model for the EB traffic. Similar to what observed from the WB traffic, both simulations tend to overestimate the utilization of the middle lane (lane 2), and underes timate the utilization of the left lane (lane 3). The only difference is that the new model tends to overestimate the utilization of the right lane

PAGE 190

190 (lane 1), while the original model tends to unde restimate this value. The overall results from the new model are closer to field observations for all three lanes. The RMSE is calculated for the vehicle lane distributions in CORSIM ori ginal model and the new lane model as 0.0571 and 0.0294 respectively, which indicate an improvement of 48.49 %. 7.3.3 Comparison of the Vehicle-Based Cumulative Number of Lane Changes The cumulative number of lane changes by vehicles as observed from the video was obtained and compared against the simulation re sults of the new mode l and the original model. As presented in Figure 7-12, CORSIM o riginal model under predicted the number of more-than-one lane changes, which is probably because CORSIM model only considers the destination, incident (including work zone) and lane use restric tions as the invoking reasons for lane changes. The other potential scenarios pr evailing on the road are not being taken into account. By incorporating the ne w scenario-based lane-changing probability model, the new model performs much better than the original model, particularly in terms of predicting the higher number of lane changes. The RMSE, as defined in Eq. 7-2, is calculated for the percentage of vehicles in CO RSIM original model and the new lane-changing model as 0.0397 and 0.0275 respectively, which indicate an improvement of 30.71%. 4 1 2)( 4 1i obs i sim iYY RMSE (7-2) where: i is the number of lane changes by vehicles, (i = 1, 2, 3 and 4), sim iY and obs iY are the observed and simulated percenta ge of vehicles with number of lane changes equaling to i.

PAGE 191

191 # of LCs by Vehicles (for Entire Arterial, WB & EB)0 0.1 0.2 0.3 0.4 1234+ Number of Lane ChangesPercentages of Vehicles New Field Data CORSIM Figure 7-12. Comparison of the cumulati ve number of lane changes by vehicles 7.3.4 Sensitivity Analysis To capture the uncertainty of how the lane -changing model affects the simulation output quantitatively, both calibrated networks (CORSIM and new) we re simulated by the O-D flows measured from different days Newberry Road vi deo data under various traffic conditions (PetArmacost et al., 1999): Off-peak period (1 hour, from 12:00 pm 1:00 pm on May 3rd, 2005); Transitional traffic period (0.5 hour, fr om 3:50 pm 4:20 pm, May 3rd, 2005); PM peak period (1 hour, from 5:00 pm 6:00 pm, April 29th, 2005). The simulation outputs include measurements fr om multiple runs for each traffic condition. The goodness-of-fit measures to eval uate the overall performance of both simulation models, in terms of average lane travel speed, average lane vehicle counts and vehicle-based number of lane changes, can be assessed numerically by standard linear statistics, such as mean error (ME), mean percent error (MPE), RMSE and root mean square percent error (RMSPE). These statistics help to quantify the overall e rror of the simulation results (Toledo and Koutsopoulos, 2004).

PAGE 192

192(1) Lane-based travel speeds from the three sections (for every 5 minutes): The lane-based average travel speeds for each section (both in WB and EB) were obtained from field data on every 5-minute interval under di fferent traffic conditions. The simulated measurements were then obtained from multiple surveillance dete ctors. The comparisons of the goodness-of-fit measures are presented in Table 7-8 (the upper pa rt). As shown, for both transitional period and peak traffic conditions, the new model outperfor ms CORSIM original model in terms of all four statistical indices. The largest improvement occurred in the simulation within the period of transitional traffic flows, dur ing which the highest percenta ge of vehicle cooperation and competition may be involved. In general, drivers during off-peak period are likely to take free lane changes, while they may either have to force merge or choose to keep on the same lane under peak congested traffic. One interesting poi nt is that under off-peak traffic condition, simulation results from the new model are not as good as that of the CORSIM original model in several indices (ME, MPE and RMSPE). This reduction of performance was believed due to the fact that many drivers during off-p eak are not likely to negotiate even though the criteria are satisfied. During particular lane-changing scenarios, vehicles may choose to accelerate or decelerate to change lane instead of negotiate with vehicles on the target lane. Consequently, the cooperative and competitive be haviors were not universally adopted in offpeak traffic. In addition, the stochastic dynamics in micro-simulation were also considered as possible sources.

PAGE 193

193Table 7-8. Goodness-of-fit stat istics for lane-based travel speeds, ve hicle counts and number of lane changes Off-Peak (60 min) Transitional (30 min) PM-Peak (60 min) Statistics Measure Original New Impro. OriginalNew Impro. Original New Impro. Average Lane-based Travel Speeds (WB & EB) ME (mph) -1.98 -2.09 -5.6% -1.79 1.02 43% -1.67 1.35 19.2% MPE (%) -4.96 -5.41 -9.1% -7.63 4.74 37.9% -8.92 6.74 24.4% RMSE (mph) 5.73 5.51 3.8% 3.14 2.53 19.4% 3.45 2.83 18% RMSPE (%) 16.42 18.12 -10.4% 12.81 9.21 28.1% 12.19 9.62 21.1% Average Lane-based Vehi cle Counts (WB & EB) ME (veh/5min) 1.15 1.24 -7.8% 1.52 1.09 28.3% 2.06 1.94 5.8% MPE (%) 6.03 5.72 5.1% 5.36 3.57 33.4% 11.24 10.26 8.7% RMSE (veh/5min) 5.26 4.83 8.2% 4.28 3.11 27.3% 4.12 3.85 6.6% RMSPE (%) 21.08 18.14 13.9% 16.43 10.02 39% 17.43 14.87 14.7% Vehicle-based number of lane changes (WB & EB) ME (veh) -0.028 0.020 27.3% -0.048 -0.014 70.5% -0.043 -0.030 29.4% MPE (%) -0.369 0.381 -3.1% -0.637 -0.308 51.7% -0.403 -0.340 15.6% RMSE (veh) 0.043 0.029 31.75% 0.063 0.040 36.7% 0.044 0.032 28.8% RMSPE (%) 0.068 0.102 -50% 0.358 0.287 19.7% 0.197 0.163 17.1%

PAGE 194

194(2) Lane-base vehicle counts from the four locations (for every 5 minutes): The field lanebased vehicle counts (both in WB and EB) were observed from the four locations as mentioned in Figure 7-2, and aggregated on every 5-minute inte rval to obtain the traffic counts on each lane. The simulated lane specific vehicle counts from the two models were obtained and compared against the actual observations under different traffic conditions. As presented in Table 7-8 (the middle part), the new model has a significan tly better match with the actual observations during the transitional traffic pe riod. For PM peak condition, the performance of the new model is also better than CORSIM original model. For the off-peak traffic, most of the statistics from the new model indicate th e improvement of performance except ME, which deceases as compared to the results from CORS IM original model. Since the new model outperformed the original model particularly in terms of transitional traffic and PM peak traffic, the reduction in ME is probably caused by the white noise points in the microsimulation instead of systematic outcomes. Consequently, it was concluded that the overall performance of the new model is bette r in matching with the field observations. (3) Cumulative vehicle-based number of lane changes (for entire simulation period): The cumulative number of lane changes for vehicles were observed from the different field videos, and aggregated to the percentages of lane-changing number equaling to 1, 2, 3 and 4+. The simulated vehicle-based number of lane cha nges from the two models were obtained and compared against the actual obs ervations under different traffic conditions. Here, an additional plug-in has been developed to record the num ber of lane changes for CORSIM model. As presented in Table 7-8 (the lower part), both models have a pretty good match with the actual observations for the off-peak traffic. For the tr ansition traffic, the performance of the new model is much better than CORSIM original m odel. All four statistics from the new model

PAGE 195

195 were improved by at least 19%, which means that the new model replicat es field observations much better. For PM peak traffic, both models te nd to underestimate the number of lane changes. This may be because under congested traffic, many other potential reasons may invoke a lane change, which depend largely on the drivers hu man behavior and mental status. Both the original and new models may not fully cap ture these invoking reasons. However, the performance of the new model is also better than CORSIM original model, which may be interpreted that the new model captured more i nvoking reasons than the original model. Since the new model outperformed the original model in terms of all three levels of traffic, it was concluded that the overall perfor mance of the new model is better in matching with the field observed vehicle-based numb er of lane changes. 7.4 Summary and Conclusions The lane-changing model was implemented and va lidated using CORSIM simulator in this chapter. With the Newberry Road (Gainesville, FL) arterial segment and the field-measured traffic and control data, the new model was ev aluated and compared to the original CORSIM lane-changing model. First, the Newberry Road (Gainesville, FL) arterial segment was simulated and calibrated with the original lane-changing algorithm in CORS IM. The calibration at this stage is to tune up CORSIM embedded sensitive behavioral parameters related to lane-changing. Next, the new lane-changing algorithm was implemented as an R TE plug-in and invoked in the simulation to replace the original lane-cha nging strategy. The new model was applied to simulate the Newberry Road (Gainesville, FL) arterial segment. The calibration endeavor of the new simulation is to tune up behavior-related parame ters within the newly developed lane-changing model. Lastly, the new lane-changing model was simulated with the additional OD demands measured from congested traffic and validated against the CORISM original lane-changing

PAGE 196

196 model. The measures of validati on include comparison of average lane-based travel time, lane distribution of vehicles, and cumulative lane ch anges by vehicles. The validation results support significant improvement over CORSIM original lane-changing model, which ignores some possible lane-changing scenarios and vehicle interactions involved. The improvement in the model performance was demonstrated through a sensitivity analysis with traffic from different times of the day. Standard statistics for comparing the goodness of fit measures (new and CORISM) i ndicate that the new lane-changing model performs best during the early afternoon (3:50 pm to 4:20 pm) traffic, which represents a moderately congested arterial situation. Large percentages of improvement on each measure of performance were found. In addition, the new model also outperforms CORSIM original model during the later afternoon peak hour (4:30 pm to 5:30 pm) traffic, even though the improvement values are not as large as those fr om the early afternoon traffic. For the non-peak hour (non-congested) traffic, the performances of th e new and original models are very close to each other. However, the validation resu lts support that the new model has better capabilities to replicate lane -changing maneuvers under modera tely congested and congested traffic in terms of various measures of performance. This, consequently, demonstrates improvement in the simulation capabilities of the new models. Validation is an essential step of model development. In this research, CORSIM (NETSIM) microscopic simulator was selected to implem ent and simulate the proposed lane-changing model. By implementing the proposed model as an RTE plug-in within CORSIM, various analyses and comparisons were conducted to prove the effectiveness of the model. The validation field data were collected from a highly congested arterial segment, Newberry Road. Unfortunately, one particular la ne-changing scenario, a work z one, was not involved, and the

PAGE 197

197 lane-changing probability functi on for the corresponding scenario was not tested and validated. Moreover, one of the DLC scenarios, changing lanes due to attracted by a better pavement condition, can not be simulated in CORSIM. Consequently, additional arterial segments and modeling techniques may be introduced for validation purposes. Additionally, CORSIM microscopic simulator is selected as the test platform in this validation. Other widely used commercial simulators, such as AIMSUN, PARAMI CS and VISSIM, may also be considered as test beds for the model validation, which would form a very important extension to the current research.

PAGE 198

198 CHAPTER 8 CONCLUSIONS This chapter summarizes the research results presented in the thesis and highlights the major contributions. Future res earch directions are provided at the end of the chapter. 8.1 Research Summary As one of the fundamental driver behaviors, the decision to change lanes depends on many factors. In this thesis, a comprehensive fram ework was presented to model drivers lanechanging behavior on arterials as a four-s tep decision-making pro cedure: decision under particular lane-changing reasons, target lane selection, gap accep tance and vehicle movement to the target lane. Emphases are placed on the first and third steps, in which any possible conditions may affect the drivers final decision. One major objective of this thesis is to model the drivers lane-changing behavior under conge sted traffic in a microscopic perspective, with special attention to the effects of driver characteristics upon the maneuvers. The lane-changing model presented in this th esis integrates a lane-changing probability component and a gap acceptance component, which capture driv ers lane-changing decision under each of the DLC reasons and different gap acceptance situations. The emphasis is to model the lane-changing maneuvers by using the driver behavior-related data along with driver background and characteristics. Tr aditional external observations based vehicle data, such as vehicle trajectories, only provide rudimentary traffic information and are not sufficient to expose the drivers thinking process duri ng the maneuver. As a result, the driver characteristics were not able to incorporate into the driver behavior research. In this study, focus group study was conducted to obtain driver behavi or related data, such as pers onal perceptions and attitudes regarding lane-changing maneuvers, which can be used to model lane changes within an urban street environment. Taking into consideration both the pers onal background data and the stated

PAGE 199

199 behavior data obtained from the discussions, the focus group participating drivers were classified into four groups using clustering analysis. The important factors for each of the lane-changing scenarios were obtained from the study, so that a further in-vehicle fiel d experiment could be designed to collect the correspond ing field lane-changing data. As the verbal responses of the focus group part icipants may not reflect their actual driving behavior, an in-vehicle fiel d experiment was conducted as fo llow-up research to test and validate the lane-changing process and the stated prefer ences from a diverse group of drivers. The quantitative values for the important factor s proposed by the focus group participants were collected from the in-vehicle field driving tests. Cluster anal ysis, similar to the one conducted for the focus group data, was then performed to categorize the participa ting drivers based on the selected measures of driver behavior. The clus tering result (into four groups) was found to be consistent with the result obtained from the focus group analysis, which confirms the effectiveness of driver classification scheme in the urban arterial lane-changing modeling. Three types of lane-changing related maneuvers (potential, attempted but unsuccessful, and completed) were collected during the in-vehicl e driving test. The combination of attempted and completed lane changes indicates the driver accepted the particular DLC scenario, while the potential maneuvers indicate a rejection in the lane-changing reason level. With the important factors obtained from the focus group study and th e quantitative values co llected from the invehicle experiment, the lane-changing probability under each of the DLC scenarios is modeled as a function of corresponding important factors and driver types. The modeling coefficients were estimated using the binary logistics (accept or not) regression method. This component tries to enumerate all DLC scenarios occurring in urban arterials and models each one individually.

PAGE 200

200 With the gap acceptance strategies observed an d the behavioral-related values collected from the in-vehicle experiment, a new lane-changing gap acceptance algorithm was developed to model lane changes on urban arterials into three modes: (i) free, (ii) forced, and (iii) competitive/cooperative. The free and forced lane changes were modeled as instantaneous events conducted during the time interval immediately following the drivers decision. The subject vehicle is moved to the target lane, and the car-following strategy is applied subsequently to the corresponding vehicles. The procedure of compet itive/cooperative lane changes is modeled as a sequence of hand-shaking nego tiations between vehicles with more complex interactions. Various strategies were develope d to model the interactions that may occur during the maneuver. The multi-agent techniques were adopted to model each vehicle as an intelligent and autonomous entity, which observes and acts upon the driving environment. The new lane-changing model was implem en ted as an individual module within the microscopic traffic simulator, CORSIM and validat ed using aggregate real -world data: the field data collected from Newberry Road (Gainesville, FL). Part of the available aggregate data was first used to calibrate the overall simulation system. The remaining aggregate data (not used for calibration) were then compar ed with the corresponding outpu ts of the calibrated CORSIM new model and those generated from the ori ginal CORISIM lane-changing model. Various goodness-of-fit measures were calculated to valid ate the improvement of the new model. The aggregate validation results demonstrated that the newly developed model performed consistently better than the or iginal CORSIM model for both moderately congested and heavy congested traffic conditions, which exhibite d the improved performance in simulation capabilities to replicate th e lane-changing behaviors.

PAGE 201

2018.2 Contributions This thesis advances the state of the art in modeling drivers lane-changing behavior on urban arterials. The major contributions are the methods used to introduce driver characteristics into the lane-changing study. Focus group survey s and in-vehicle observations were conducted to collect microscopic data for the modeling purp ose. The in-vehicle field data were used to estimate the lane-changing probability models by statistically rigorous methods (regression analysis). The developed lane-changing model ha s bridged some of the significant gaps in the existing simulation tools. The specific contributions of each empirical study are listed below: (1) Contribution to the modeling framework on driver characteristics: In this research, two important components within the lane-chang ing behavior, reason-based lane-changing probability and gap acceptance, are studied. The ex isting lane-changing models either borrow the reasons and the acceptable gaps from other models, or extract them from the given video data (Hidas, 2002, 2005; Liu et al., 2006), or esti mate them from field data (Ahmed, 1999; Choudhury et al., 2004; Ben-Akiva et al., 2006). No one considers driver characteristics in much detail, since the information is hardly be deduced from the exis ting videos or other sources of field data. This research was st arted with the focus group study, and thus the factor of driver characteristics was incorporated into the la ne-changing behavior data collection from the beginning. By using both personal background and th e stated behavior data related to urban arterial lane changes, the focu s group participants were categori zed into four groups. Then with the in-vehicle data collected from field lane -changing maneuvers, a further cluster analysis was conduct to classify the in-vehicle drivers. The result was found to be consistent with the one obtained from the focus group study, which in turn validates and confirms the output from the focus group study. With the field-collected lane-changing values and the corresponding driver types, a comprehensive model was developed to handle the probability of changing lanes

PAGE 202

202 under each particular scenario and the strategi es adopted in the subsequent gap acceptance procedure. (2) Contribution to the empirical work of data collection: Driver behavior is largely dependent on personal characteri stics and driving expe rience. To obtain representative lanechanging data, three types of data collections were included in this research. Driving experience based focus group discussions were first conducted. A diverse pool of participants was recruited based on age, gender, driving experience, occu pation and vehicle owners hip. Driver background and the qualitative lane-changi ng data were collected from the focus group study. Next, according to the results of the focus group study, an in-vehicle experiment was designed to collect field quantitative values The testing drivers were acco mpanied by the researcher on the pre-selected route to collect th e behavior data related to comp leted, attempted but unsuccessful, and potential lane changes. Additional cluste ring analysis was conducted to validate the consistency of the in-vehicle data with the focus group results. Finally, in addition to the focus group study and the in-vehicle data collection, external observati on based field video data were also collected and used for model calibration and validation purposes. (3) Contribution to model vehicle interactions during lane changing: Three lane-changing modes were identified based on vehicle interactions as: free, forced and cooperative/competitive lane changes. The procedure of the competitive/ cooperative lane changes is modeled based on drivers actions and responses as a sequence of hand-shaking negotiations, by referring to the protocols in computer network communications. Strategies were developed according to the immediate surrounding environment and the correspondi ng driver characteristic s, so that vehicle actions can be modeled correctly. This approach differs from existing models which assume that lane changes are all conducted instantaneously or within fixed time intervals, and different lane-

PAGE 203

203 changing modes are not interchangeable. In the new gap acceptance model, the games between the subject vehicle and the lag vehicle may be either competition or cooperation, depending on the surrounding traffic and the characteristics of drivers. The stra tegies of not change, free change, cooperative/competitive change and forced change are interchangeable, which better reflects the lane-changing reality on urban arterials. (4) Contribution to model implementation and validation: The proposed lane-changing model was implemented as a dll in VS .net C ++ 2008, and validated in CORSIM microscopic simulator. First, the individual component, in cluding the probability functions for the lanechanging reasons and the gap acceptance procedure, is implemented as a separate function (or called subroutine) within the lane-changing module, which is invoked as a CORSIM RTE (run time extension) during the simulation. Next, tw o CORSIM simulation cases, the one with the newly developed lane-changing model and the one with CORSIMs original lane-changing model, were calibrated using aggregate data (f ield video data) collected from arterials in Gainesville, FL. Once the calibrations were comple ted, values of the full set of behavioral parameters were fixed. Both calibrated models are simulated with the additional set of OD demands for validation purpose. The results ar e compared with the field measurements on multiple indices of the measures. Various goodness-of-fit statistics were generated to determine the agreement between the results from the simu lation system and the field observations. The simulation results from the existing lane-changing model in CORSIM served as a test bed for the model ability improvement offered by the new al gorithm. The analyses of results indicated, by incorporating the new lane-chang ing model into micro-simulation tools, more realistic traffic flow and congestion can be represented and assess ed, which hence results in better planning and policy analysis tools.

PAGE 204

2048.3 Directions for Future Research In this thesis, a general framework for modeli ng lane-changing behavior on urban arterials was presented. The focus group study and in-ve hicle experiment we re conducted to obtain information on driver characteristics, which we re then incorporated into the lane-changing behavior modeling. The research concept and th e proposed framework have enormous potential both in modeling driving decisions and modeling decisions in other scenarios. Some of the directions in which furt her research is needed are presented below: (1) Target lane selection component: The lane-changing process is generally modeled as a sequence of four decision-making steps: accep tance of lane-changing reasons, target lane selection, existing gap acceptanc e and vehicle movement to targ et lane. The two highlighted components developed in this thesis are th e lane-changing probabil ity model and the gap acceptance model. The target lane was assumed to be always the adjacent lane, which is generally referred to as a myopic target lane model. This assumption does make sense for modeling the regular urban arterial traffic. Ho wever, it can not be applied to the special situations on urban arterials, such as the ex istence of BRT (Bus Rapid Transit) or HOV (High Occupancy Vehicle) lanes, which require additional considerations. To this end, a target lane choice model can be adopted to evaluate the utilitie s of all candidate lanes, so that the lane with highest utility is chosen as the ta rget lane. Variables likely to influence the target lane choice of the driver include path-plan, lane attributes driving style and capab ilities, and so on. (2) Additional DLC scenarios and the corresponding important factors: One highlight of the research results is the scenario-based lane-changing probability model, which models the probability of changing lanes under each pre-selected DLCs as a function of indicated important factors and the subject driver type. The important factors for each DLC were obtained from the

PAGE 205

205 focus group study and are not inclusively applic able to all other pervasive DLCs on urban arterials. Consequently, the current model can on ly handle the scenarios discussed in the focus group meetings. More focus group discussions are required for incorporating new DLCs into the probability model. (3) Vehicle type effect: During the focus group meetings, al l lane-changing related questions and scenarios assumed that the subject vehicle wa s a passenger vehicle. Furthermore, the subject vehicle was fixed as the Honda Pilot throughout the in-vehicle experiment. Consequently, the proposed model only captures the lane-changing be havior of passenger vehicles. As revealed during the focus group discussions, the lane-changing behavior of h eavy vehicles differs largely from that of passenger cars, which is an interestin g direction to extend this research in the future. (4) Other promising data collection technologies: One significant component in driver behavior modeling is the data collection, which is especially important in the lane-changing behavior research. In this thes is, two methods, video and film me thods and instrumented vehicles (with GPS systems), were used to collect field data on typical urban arte rial segments. Further studies may consider technologies for collect ing data from more versatile geometric characteristics. To this end, virtual reality driv ing simulators may be used to collect data in situations that are otherwise di fficult to observe, such as emer gency situations, and to control some of the latency in the behavior (e.g. by aski ng drivers to perform a specific maneuver, thus eliminating the uncertainty in modeling the drivers short-term goals).

PAGE 206

206 APPENDIX A MEMORANDUM OF THE SUBMIT MATERIALS CHECKLIST MEMORANDUM (UFI RB #2008-U-0019) DATE: January 8th, 2008 TO: UF IRB-02 (UF Campus/Non Medical) FROM: 1. Daniel (Jian) Sun, Graduate student, Department of Civil and Coastal Engineering jiansun@ufl.edu 2. Lily Elefteriadou, Ph.D., Associate professor Department of Civil and Coastal Engineering elefter@ufl.edu SUBJ.: UFIRB Subm ission Checklist Survey of the Lane-Changing Behavior in Urban Arterials In accordance with the requirements of the Univer sity of Floridas Institutional Review Board (IRB), all research involving human subjects ne eds to be approved by the relevant IRB Office prior to conducting any activities. This documen t lists the materials hereby submitted to IRB-02 (UF Campus/Non Medical) for the pr oject titled survey of lane -changing behavior in urban arterials. The project constitutes Mr. Daniel (J ian) Suns dissertation supervised by Dr. Lily Elefteriadou in the department of Civil a nd Coastal Engineering, wherein two types of TRANSPORTATION RESEARCH CENTER (TRC) UNIVERSITY OF FLORIDA 518A Weil Hall, P.O. Box 116580

PAGE 207

207 experiments focus group discussi on and in-vehicle field data collection are proposed. The items attached are: One copy of the UFIRB protocol form (w ith original signatures, APPENDIX B), Three copies of the two informed consent fo rms (one form for each experiment, APPENDIX C and APPENDIX D), Advertisement for participants recruitment (APPENDIX E), One copy of the complete methods section from the proposal of dissertat ion research (Chapter 3), and Other research instruments, including 1. One copy of the prescreening questionnair e for participant selection (APPENDIX F), 2. One copy of the driving background survey questionnaire to be used during the participants check-in procedure of the both experiment s (APPENDIX G), 3. One copy of the script to be used during the focus group discussion (APPENDIX H), 4. Description of the instrumented vehicle wh ich will be used by participants in the field data collection (Chapter 3), and 5. Proposed routes that the study participants will drive duri ng the field data collection (Chapter 5).

PAGE 208

208 APPENDIX B UFIRB PROTOCOL FORM UFIRB 02 Social & Behavioral Research Protocol Submission Title of Protocol: Survey of Lane-Changing Driving Behavior in Urban Arterials Principal Investigator: Daniel (Jian) Sun UFID #: 8068-6760 Degree / Title: Graduate Student Department: Civil and Coastal Engineering Mailing Address: 518A We il Hall, PO Box 116580 Email Address & Telephone Number: jiansun@ufl.edu / (352)682-8390 Co-Investigator(s): UFID#: Supervisor: Lily Elefteriadou UFID#: 1319-1914 Degree / Title: PH.D, Associate Professor Department: Civil and Coastal Engineering Mailing Address: 512 Weil Hall, PO Box 116580 Email Address & Telephone Number: elefter@ce.ufl.edu / 392-9 537 ext 1452 Date of Proposed Research: Jan. 2008 to Jan. 2009 Source of Funding N/A Scientific Purpose of the Study: To capture the significant factors wh ich affect lane-changing maneuvers for different types of drivers, so that lane-changing dr iving behavior can be modeled in a more realistic way considering individual driver behavior. Describe the Research Methodology in Non-Technical Language: ( Explain what will be done with or to the research participant. ) In this research, participants will be recruited for two experiments to collect lane-changing behavior data. In the first experiment, participants w ill join a focus group to discuss lane-changing related questions from their personal driving experiences. Nex t, participants, not necessary the same as the ones involved in the focus groups, will drive an instrumented vehicle along prespecified routes to collect personal lane-changing behavior data. A background survey, which contains questions regarding age, gender, driving experience, nationality, occupation and vehicle ownership, will be conducted for each driver, so that the observed lane-changing behavior could be connected with particular drivers characteristics in t he further research stages. Describe Potential Benefits and Anticipated Risks: ( If risk of physical, psychological or economic harm may be involved, describe the steps taken to protect participant.) The survey will help model lane-changing behavior in a more realistic and accurate way. Ba sed on this fundamental component, the performance of various traffic operation models will be improved. The surv ey method adopted in this research provided a new

PAGE 209

209 methodology for the driver-oriented research in traffic engineering. No risk is anticipated during the focus group discussion experiment. For the field data collection, sinc e all drivers will be accompanied by the researcher, and be instructed to drive as they usually do, the risks will be those a driver usually as sumes during driving. The consent form includes language regarding potential injury during the experiment. Describe How Participant(s) Will Be Recruited, the Number and AGE of the Participants, and Proposed Compensation: A diverse pool of participants will be selected based on age, gender, driving experience, nationality, occupation and vehicle ownership. A prescreening questionnaire has been developed to help identify qualified participants. The questions will be posted on the project website (http://grove.ufl.edu/~jiansun). Re spondents can choose to submit responses online or download the prescreening questionnaire from the server, and submit responses offline through email or mail. Advertisements for recr uitment will be announced through various publications and several list servers. The proposed public locations for pos ting the announcement include the University of Florida campus, downtown transit transfer station, Alachua county library and supermarkets. In addition, the advertisement will be placed on the Classifieds in A lligator, and sent to several large organizations and communities, such as ASCE, FACSS and ISA, through thei r list servers. The advertisement will also be posted on the researchers personal website. Criteria for the participant recruitment are as follows: 1. mu st be a regular driver with a driving experience no less than three years; 2. must be a Gainesville resident for more than 1 year; and 3. must indicate a willingness to join either the focus group or agree to test drive the instrum ented vehicle, or both. In the focus group studies, three groups will be recruited with 5-7 participants each. The di scussion time is set as two hours, and the compensation is $50 per participant. For the in-vehicle fi eld data collection, the number of participants is set as 30-40, and the compensation is $50 each. Describe the Informed Consent Process. Include a Copy of the Informed Consent Document: The recruiting advertisement explains the purpose of the project and the objectives of the research. It is clearly stated that participation is optional and that the outcome w ill be summarized in a manner that does not identify any participant. A separate copy of the informed consent document attached will be used to advise potential participants and obtain voluntary agreement at the beginning of the experiments. Principal Investigator(s) Signature: Supervisor Signature: Department Chair/Center Director Signature: Date:

PAGE 210

210 APPENDIX C INFORMED CONSENT FORM FOCUS GR OUP STUDY Protocol Title: Focus Group Survey of Lane-Changing Driving Behavior in Urban Arterials Please read this consent document carefully befo re you decide to partic ipate in this study. Purpose of the research study: The purpose of this study is to capture the si gnificant factors which affect lane-changing maneuvers for different types of drivers, so th at lane-changing driving be havior can be modeled in a more realistic way. What you will be asked to do in the study: You will join another 4-6 volunteers to form a focus group. Then a moderator will ask you questions related to lane-changi ng behavior, and you will be require d to think and answer these questions from your personal driving experience During the session, the moderator will first present a list of possible lane -changing scenarios, and ask you to evaluate the likelihood you would change lanes for those reasons. Then the moderator will ask you to describe how you execute such maneuvers. During the experiment, you will be encouraged to interact with the other participants in your group, thereby providing greater insight into why certain beliefs and opinions are held. You will not be required to vote or reach consensus. With your permission, I would like to tape record the discussion so that I can more accurately record your responses after we finish today. Only research ers involved in this project will have access to the tape. Your identity will not be revealed in the final manuscript. The time re quired for this activity is about two hours. Risks and Benefits: No risk is anticipated during the focus group discussion experiment, and we do not anticipate that you will benefit directly by pa rticipating in this experiment. Compensation: You will be paid $50 compensation for participating in the focus groups experiment of two hours. Confidentiality:

PAGE 211

211 Information collected from this experiment will be used for traffic engineering research only. Your identity will be kept confidential to the extent provided by law. In accordance with the Confidential Information Protection and Statistical Efficiency Act of 2002 (Title 5 of Public Law 107-347) and other applicable Federal laws, your re sponses will not be disclosed in identifiable form without your consent. Voluntary participation: Your participation in this study is completely voluntary. There is no penalty for not participating. Right to withdraw from the study: You have the right to withdraw from the focus group discussion at anytime without consequence. The compensation will be recalculated based on y our participating time. If you withdraw after one hour of discussion, you will be paid $20. No compensation will be paid if the participating time is less than one hour. Whom to contact if you have questions about the study: Daniel(Jian) Sun, Graduate Student, Department of Civil and Coast Engineering, Room 518, Weil hall, (352)682-8390. Lily Elefteriadou, Ph.D., Department of Civil and Coast Engineering, Room 512, Weil hall, (352)392-9537 x1452. Whom to contact about your rights as a research participant in the study: UFIRB Office, Box 112250, University of Flor ida, Gainesville, FL 32611-2250; ph 392-0433. Agreement: I have read the procedure described above. I volunt arily agree to pa rticipate in th e procedure and I have received a copy of this description. Participant: _______________________________________ Date: _________________ Principal Investigat or: _______________________________ Date: _________________

PAGE 212

212 APPENDIX D INFORMED CONSENT FORM IN-VEHICLE EXPER IMENT Protocol Title: In-vehicle Data Collection Survey for Lane-Changing Driving Behavior in Urban Arterials Please read this consent document carefully befo re you decide to partic ipate in this study. Purpose of the research study: The purpose of this study is to collect the si gnificant factors which affect lane-changing maneuvers for different types of drivers, so that lane-changing drivi ng behavior could be modeled in a more realistic way. What you will be asked to do in the study: In this experiment, you will be accompanied by our researcher to drive in an instrumented vehicle (Honda Pilot) for 40-60 minutes. Before starting the vehicle, th e check-in procedure will be as follows: 1) Sign the informed consent form (this form); 2) Finish the background survey form; this form contains questions regarding you r age, gender, driving experience, nationality, occupation and vehicle ownership; 3) Provide you r drivers license for authentication; and 4) Turn off your cell phone, if you have one. You will be shown the map of a pre-selected route. Please review the route carefully, and try to clarify any questions you may have. If during driving you make the wrong turn for more than 3 times, the in-vehicle data collection will have to be terminated. A camera will be monitoring your face movement throughout the test. During the experiment, please follow the researchers instru ctions as closely as po ssible. To avoid driver distraction, the researcher w ill not be communicating with you while you are driving. However, you will be instructed to stop at specific check points, and spend approximately 3-5 minutes to communicate with the researcher, so that information from each driving stage can be obtained. For a successful lane change, you may be asked to give out the reasons for the maneuver, and the major factors during the lane change. For an unsuccessful one, you may need to explain the reasons you felt it was unsuccessful. The field data collected in the experiment, including both the vehicle trajectory and the information obtained during your interactions with the researcher, will be used for traffic engineering research only. With the information from the background form, your driving behavior will be connected with some of your characteristics. Your identity will not be revealed in the final manuscript. The time planned for this activity is 40 minutes driving and an additional 10-20 minutes communicating with the researcher. Risks and Benefits: The risks for this experiment are those a driver usually assumes during driving. Please be fully attentive to your driving, and have safety as you r first priority. There is an increased risk of accidents when the driver is distracted by other activities, such as talking on the phone. You will be accompanied by the principal investigator, and you will be instructed to drive as you usually do during the data collection. We do not anticipate that you will benefit di rectly by participating in this experiment.

PAGE 213

213Compensation: You will be paid $50 compensation for participating in this field data collection experiment. No compensation will be paid for withdrawing early making the wrong turn for more than 3 times or being involved in a road accident. Research-related injury: In the event that this research activity results in an injury, treatment will be available, including first aid, emergency treatment and follow-up care as needed. Care for such injuries will be billed in the ordinary manner to you or your insurance company. If you think that you have suffered a research related injury, let the study researcher know right away. Confidentiality: Information collected from this experiment will be used for traffic engineering research only. Your identity will be kept confidential to the ex tent provided by law. In accordance with the Confidential Information Protection and Statistical Efficiency Act of 2002 (Title 5 of Public Law 107-347) and other applicable Federal laws, your re sponses will not be disclosed in identifiable form without your consent. Voluntary participation: Your participation in this study is completely voluntary. There is no penalty for not participating. Right to withdraw from the study: You have the right to withdraw from th e study at anytime without consequence. Whom to contact if you have questions about the study: Daniel(Jian) Sun, Graduate Student, Department of Civil and Coast Engineering, Room 518, Weil hall, (352)682-8390. Lily Elefteriadou, Ph.D., Department of Civil and Coast Engineering, Room 512, Weil hall, (352)392-9537 x1452. Whom to contact about your rights as a research participant in the study: UFIRB Office, Box 112250, University of Flor ida, Gainesville, FL 32611-2250; ph 392-0433. Agreement: I have read the procedure described above. I volunt arily agree to pa rticipate in th e procedure and I have received a copy of this description. Participant: _______________________________________ Date: _________________ Principal Investigat or: _______________________________ Date: _________________

PAGE 214

214 APPENDIX E ADVERTISEMENT FLYER FOR THE PARTICIPANTS RECRUITMENT

PAGE 215

215 APPENDIX F PRESCREENING QUESTIONAIRE FOR PARTICIPANTS SELECTION

PAGE 216

216

PAGE 217

217 APPENDIX G PARTICIPANTS DRIVING BACKGROU ND SURVEY QUESTIONAIRE

PAGE 218

218 APPENDIX H FOCUS GROUP MODERATIING SCRIPTS Thank you for taking time out of your busy schedule to be here today. Daniel Sun (under advising of Dr. Lily Elefteriado) is interested in underst anding how drivers perceive the lane-changing behavior for his ph.d studies, which includes why drivers change lanes and how they behave during the maneuver. As you may know, multiple surveys have been conducted as useful data collection methods by professors in UF TRC in the past several years. However, focus groups such as this one have the unique advantage of gathering information without constraining participants to a predetermined set of responses. I will be asking you as a group a number of questions over the next two hours. I would like you to be completely honest with me, and answer questions from your own drive experience. I assure you that all of your responses will be held in complete con fidence. I would like to ask your permission to tape record the whole discussion so that I can more accurately study your responses after we finish today. I will not link you with any of your comments after the discussion, and no identifying information will leave this room. Are all of you comfortable with this? Opening question (I I c c e e b b r r e e a a k k e e r r q q u u e e s s t t i i o o n n, which will get everyone to talk and help participants to feel comfortable.) 1. Let the participants introduce themselves to one another. Open th e discussion with the following question Tell us who you are. Do you enjoy driving? Why? (3-4 minutes) Introductory question (Introduce the topic of discussion and get people to start thinking about their connection with the topic.) 2. What comes to your mind when you hear the term change la nes? (5-7 minutes) Transition question (Move the conversation into the key questions.) 3. Think about the reasons invoking a lane chan ge. Do you consider there are many differences between you and other drivers? (5-8 minutes) Key questions (Key questions drive the study) 4. In general, lane-changing behavior could be divided into mandatory lane changes (MLC) and discretionary lane changes (DLC). MLCs are th ose initiated to follow a special route. The purpose of DLC is for drivers to im prove their position in the traffic stream to expedite their trip. Not all drivers make DLC in given a certain situati on. In this question, you will be given a list of DLC types. Please evaluate each one and sele ct the frequency that you would conduct such a lane change from your own dr iving experience. (20 minutes) The given list of DLCs (Discretionary lane-changing) is: R1) Passing a stopped-bus at bus stop; R2) Giving way to a merging vehicle or to a bus merging from a bus pull-off; R3) Gaining speed advantage by overt aking a slower moving vehicle;

PAGE 219

219 R4) Gaining queue advantage; R5) Avoiding a heavy vehicles influence; R6) Avoiding the closed-following pressure impose d by the vehicle behind you (only applied to the lane-changing maneuver to curb-side); R7) Attracted by a better paveme nt condition, such as away o ff snowy/icy pavement lane Five levels of frequency are as: Level 1: generally do not conduct (< 10%) Level 2: sometimes conduct but more likely do not (10% 40%) Level 3: sometimes conduct, sometimes not. It hard to conclude which one is preferred (40% 60%) Level 4: sometimes do not conduct but more likely conduct (60% 90%) Level 5: generally conduct (> 90%) Table 1 as follow will be used to collect the frequency for each reason from the participants. Table 1: Survey form used to collect the acceptable extent for each reason Reason Fre. R1 R2 R3 R4 R5 R6 R7 R8 Level 1 Y Y Y Level 2 Y Y Level 3 Y Y Level 4 Level 5 5. In addition to the list of DL Cs given in question 4, the li st of MLCs (Mandatory LaneChanging) is given as: 1) Making turning (left/right) move ment at the immediate/next downstream intersection; 2) Avoiding an incident/permanent obstruction (e.g parked vehicles because of accidents or other emergencies, or work zone lane closure,); 3) Avoiding the end of current lane Do you think besides above MLC and DLC reasons, are there any other reasons could also be account for your behavior of cha nging lanes (please enumerate them and give out the level of frequency accordingly)? (5 minutes) 6. Ill give out a scenario for each particular lane change, could you enumerate the major factors that affect y our decision to attempt a lane chan ge for each of those scenarios? Try to explain how much importance you would assign to applicable factors for each scenario? (45 minutes) Five levels of importance are as: Level 1: very important, no other factors are more important than this Level 2: important, but very important Level 3: not so important, at middle level

PAGE 220

220 Level 4: this factor will be consid ered, but it is not an important one You may be asked to explain briefly how so me acceptable DLCs are executed, and how the given MLCs are executed. Scenario 1): 6.1.a) You are approaching an intersection where a left turn will be made. Suppose you are not in the correct lane to execute the turn and need at least one lane change to the correct lane. When (how far do you from the intersection) and how will you consider a lane change? (Please describe your concerns and possible mane uvers during the process, 5-6 minutes) Left Turn Scenario 6.1.b) You are approaching an intersection where a right turn will be made. Suppose you are not driving on a lane for the turn and need at least one lane change to the correct lane. When and how will you consider a lane change? Is there any difference between this and the above situation? (Please describe your concerns and possible maneuvers during the process, 5-6min) Right Turn Scenario Scenario 2): 6.2.a) You are approaching a location of parked vehi cles blocking your lane ahead because of an incident or other specific emergencies. When and how will you consider a lane change to the

PAGE 221

221 open lane to avoid the lane obs truction? (Please desc ribe your concerns and possible maneuvers during the process) Parked vehicles because of incident or other emergencies Incident or other emergencies 6.2.b) You are approaching a location of a work zone, and the current lane is closed. When and how will you consider a lane change to the open lane to avoid the lane obstruction? (Please describe your concerns and possible maneuvers during the process) Work zone Location of workzone Scenario 3): 6.3) If you see a sign that the lane you are in will end in 1,000 ft, when and how will you consider a lane change to merge into the adja cent lane? (Please descri be your concerns and possible maneuvers during the process, 5-6min)

PAGE 222

222 Scenario 4): 6.4) When driving on your current lane, you find a bus in front is loading or unloading passengers. Will you consider a la ne change? If yes, when a nd how will you consider a lane change to pass the bus? (Please describe your concerns and possible maneuvers during the process, 5-6min) bus stopped at a bus-stop Passing a stoppedbus at bus stop Scenario 5): 6.5) You find a merging vehicle or a bus from a bus pull-off is entering into your lane. Will you consider a lane change? If yes, when and how will you consider a lane change to pass the vehicle/bus? (Please describe your concerns and possible maneuvers during the process, 5-6min)

PAGE 223

223 Scenario 6): 6.6) You find the vehicle in front of you is driv ing slower than you would like speed. Will you consider a lane change? If yes, when and how w ill you consider a lane change to pass the slow vehicle? (Please describe your concerns and pos sible maneuvers during the process, 5-6min) Slow vehicle Overtaking a slower moving vehicle

PAGE 224

224Scenario 7): 6.7) When approaching an intersection, you find the queuing vehicles in your current lane are much longer than that of other lanes. Will you cons ider a lane change? If yes, when and how will you consider a lane change to gain the queue advantage without othe r negative influences? (Please describe your concerns and possible maneuvers during the process, 5-6min) Lane-changing to gain queue advantage Queue advantage Scenario 8): 6.8) When driving on the current lane, you find a heavy vehicle in front influences your driving state. Will you consider a lane change? If yes, when and how will you consider a lane change to avoid the heavy vehicles influe nce? (Please describe your co ncerns and possible maneuvers during the process 5-6min) Scenario 9): 6.9) When driving on the center lane, you find you ar e tailgating by the vehicle behind you. Will you consider a lane change? If yes, when and how will you consider a lane change to avoid the pressure from behind? (Please de scribe your concerns and possible maneuvers during the process, 5-6 min)

PAGE 225

225 7. Now we want to capture in teractions among drivers during a lane change maneuver. Please describe your actions during a lane change ma neuver assuming the traffic is congested. Two lane-changing scenarios including both merging to the curb-side lane and the median-side lane will be investigated. Answers from this questi on would help to indicate the cooperation and competition among drivers for lane changing during c ongested traffic, so that new model may be developed for the interesting behaviors. (Note: 1. Considering the actions you will adopt if you are a merging driver or lag driver in the target lane respectively? 2. How is the gap acceptance difference between the congested traffic and nor mal traffic for each scenario?) (25 minutes) Scenario 1: Scenario 2:

PAGE 226

226 Ending questions (Intent to close the discussion, enable participants to reflect on previous comments.) 8. Today, we began with the major possible r easons that would invoke a lane change and the procedure for executing it. Then for each reason, th e major effective factors which affect drivers decision on lane change were enumerated and examined. Fina lly, the possible interactions involved in a lane change behavior were discusse d. Did I correctly describe what was said here? Is there anything you want to say but di dnt get a chance? (4-5 minutes)

PAGE 227

227 APPENDIX I K-MEAN ALGORITHM USED TO OBTAIN THE CENTROIDS Data: x(j) Nj : driver aggressivene ss for each participant j ; K : number of cluster; iu ] ,1[ Ki : centroid value for cluster i ; {M( i )}, ] ,1[ Ki : cluster set for each cluster i ; initialization: {0 1u ,0 2u 0iu } { K N ix *)1(1 }, Ki ; while ({m iu } {1 m iu }) do ; centroid for each cl uster moves (converge) for each participant x(j) do x(j) M( i ) arg min(m iu )2))((m iujx ; find the nearest cluster center and assign it to the cluster; end x(j) {m iu } min2))((m iujx ; for each cluster M( i ) do 1 m iu = avg (M( i )); recompute the centroid i for each cluster; end m = m + 1 end

PAGE 228

228 LIST OF REFERENCES AAA Foundation for Traffic Safety. A re you an aggressive driver? http://www.aaafoundation.org/quizzes /index.cfm ?button=aggressive Dec. 2006. AAA Foundation for Traffic Safety. Aggre ssive driving: research update. http://www.aaafoundation.org/pdf/AggressiveDrivingResearchUpdate2009.pdf Apr. 2009. Ahmed, K.I. (1999). Modeling drivers accele ration and lane changi ng behavior. Ph.D. Dissertation, Department of Civil and Environmental Engineering, MIT Ahmed K.I., Ben-Akiva, M.E., Koutsopoulos, H. N., and Mishalani, R.G. (1996). Models of freeway lane changing and gap acceptance behavior. Proceedings of the 13th International Symposium on Transportation and Traffic Flow Theory Lyon, France, 501-515. Albert, A. and Anderson, J.A. (1984). On the existence of maximum likelihood estimates in logistic regression m odels. Biometrika 71(1), pp. 1. Allison, P.D. (1999). L ogistic regression us ing SAS: Theory and application (ed. 1st). SAS Publishing. Badra, N.M. (2007). Sensitivity anal ysis of transportation problems. Journal of Applied Sciences Research, 3(8), pp. 668-675. Barcelo, J., Casas, J., Codina, E., Fernandez, A. Ferrer, J.L., Garcia, D., and Grau, R. (1996). PETRI: A parallel environment for a real-time traffic management and information system. Proceedings of the 3rd World Congress on ITS Orlando, USA. Barcelo, J., Casas, J., Ferrer, J. L., and Garcia, D. (1998). Modeli ng advanced transport telem etric applications with microsc opic simulations: The case of Aimsun2. 10th European Simulation S ymposium Proceedings Nottingham, England, pp. 362-367. Ben-Akiva, M.E, Choudhury, C.F., Lee, G., Rao, A., and Toledo, T. (2006). Verification and validation plan: Forced lane change and c ooperative merging model (with slides). Ngsim Group, MIT. Ben-Alive M.E. (1973). Structu re of passenger travel de mand models. Ph.D. Thesis, Department of Civil and Envi ronmental Engineering, MIT. Ben-Akiva, M.E. and Lerman, S. (1985). Discrete choice an alysis: Theory and application to travel dem and. The MIT Press, Boston. Ben-Akiva, M.E., and Bierlaire, M. (2003) D iscrete choice models w ith applications to departure time and route choice. in Hall R. (ed.) Handbook of Transportation Science Kluwer.

PAGE 229

229 Bhat, C.R. (1997). Covariance hete rogeneity in nested Logit mode ls: Econometric structure and application to intercity travel. Transportation Research Part B 31 (1), pp. 11-21. Brackstone, M., and McDonald, M. (1996). The microscopic modeling of traffic flow: Weaknesses and potential developments. In: Wo lf, D. E., Schreckenberg, M., Bachem, A. (Eds.), Traffic and Granular Flow World Scientific, Singapore, pp. 151-165. Brackstone, M., McDonald, M., and Wu, J. ( 1998). Lane changing on the motorway: Factors effecting its occurrence, and their implications. Proc. of the 9th Int. Conf. on Road Transport Information and Control London, UK, pp. 160-164 Brackstone, M., and McDonald, M. (1999). Car-following: A historical review. Transportation Research 2F pp. 181-196. Brackstone, M., Waterson, B., and McDonald, M. (2009). Determinants of following headway in congested traffic. Transportation Research 12F pp. 131-142. Brownstone, D., Bunch b, D.S., and Train c, K. (2000). Joint mixed L ogit m odels of stated and revealed preferences for alternative-fuel vehicles. Transportation Research Part B 34 (5), pp. 315-338. Cameron, G.D.B., and Duncan, G.L.D. (1996). PARAMICS Parallel microscopic simulation of road traffic. The Journal of Supercomputing 10, pp. 25-53. Carrasco, J.A., and Ortuzar, J.D. (2002). Review and assessment of the nested Logit model. Transport R eviews: A Transnational Transdisciplinary Journal 22(2), pp. 197-218. Chandler, F. E., Herman, R. and Montroll, E. W. (1958). Traffic dynamics: Studies in car following. Operations Research 6, pp. 165-184. Chien, S.I., Mouskos, K.C., and Chowdhury, S.M. (2001). Generating driver population for the microscopic simulation model (CORSIM). Simulation 76(1), pp. 40-45 Chrysler, S.T., Carlson, P., Zimmer, D., a nd Williams, A. (2004). Assessing the desired functionality and costs of an instrumented vehicle for TTI research purposes TEXCar transportation experiment car. Final Report, TTI. Choudhury, C.F. (2005). Modeling lane -changing behavior in presen ce of exclusive lanes. M.S. Thesis, Department of Civil and Environmental Engineering, MIT. Choudhury, C.F., Toledo, T., and Ben-Akiva, M.E. (2004). NGSIM draft report for lane selection model. NGSIM Group MIT. Choudhury, C.F. (2007). Modeling driving decisi ons with latent plans. Ph.D. Thesis, Department of Civil and Envi ronmental Engineering, MIT.

PAGE 230

230 Crane, S. S. (1996). An empirical study of alte rnative fuel vehicle c hoice by commercial fleets: lessons in transportation choices, cost effici ency, and public agencies organization. Ph.D., Dissertation, Department of Economics, UC Irvine. Das, S., Bowles, B. A., Zhang, Y., Houghland, C. R., and Hunn, S. J. (1999). An autonomous agent model of highway driver behavior. Transportation Research Board 78th Annual Meeting Washington, DC. Daganzo, C.F. (2006). In traffic flow, cellular automata = kinematic waves. Transportation Research 40B pp. 396-403. Duan, J. (2006). Similarity analysis for estima tion of an activity-based travel demand model. Ph.D. Dissertation, Department of Civil a nd Environmental Engineering, UC Irvine. Fellendorf, M. (1994). VISSIM: A microscopic si mulation tool to evaluate actuated signal control including bus priority. 64th ITE Annual Meeting Session 32, Dallas. FHWA (1998). CORSIM user manual (1.04 ed.) Federal Highway Administration, U.S. Department of Transporta tion, McLean, Virginia. FHWA (2006). CORSIM run-time extension (R TE) developers guide (6.0 ed). Federal Highway Administration, U.S. Department of Transportation, McLean, Virginia. Gazis, D., Herman, R., and Rothery, R. (1961). Nonlinear follow-the-leader models of traffic flow. Operations Research 9 pp. 545-567. Gipps, P.G. (1981). A behavioral car-fo llowing model for computer simulation. Transportation Research 15B pp. 105-111. Gipps, P.G. (1986). A Model for the St ructure of Lane-Changing Decisions. Transportation Research Part 20B 403-414. Gipps, P.G., and Wilson, B.G. (1980). MULTSIM: A computer package for simulating multilane traffic flows. Proc. Fourth Biennial Conference, Simulation Society of Australia pp. 88-93. Golob, T.F., and McNally, M.G. (1997). A model of activity part icipation and travel interactions between household heads. Transportation Research 31B 177-194. Gopinath, D.A. (1995). Modeling he terogeneity in discrete choi ce processes: Application to travel demand. Ph.D. Thesis, Department of Civil and Environmental Engineering, MIT. Halati, A., Lieu, H., and Walker, S. (1997). COR SIM corridor traffic simulation model. In: Proceedings ASCE Conference on Traffic C ongestion and Safety in the 21st Century Chicago.

PAGE 231

231 Hardy M.A. (1993). Regression with dummy variables(ed. 3rd). Sage Publications, Inc., Thousand Oaks, CA. Herman, R., Montroll, W., Potts, R.B., and Roth ery, R.W. (1959). Traffic dynamics: Analysis of stability in car following. Operations Research 7, pp. 86-106. Herman, R., and Rothery, R.W. (1969). Frequency and amplitude dependence of disturbances in a traffic stream. Proceedings of 4th International Sym posium on the Theory of Traffic Flow Ed. W. Leutzbach and P. Baron. Bonn, Germany. Hidas, P. (1998). A car following model for urban traffic simulation. Traffic Engineering and Control 39 (5), 300-309. Hidas, P. (2002). Modeling lane changing and merging in microscopic traffic simulation, Transportation Research 10C pp. 351-371. Hidas, P., and Wagner, P. (2004). Review of da ta collection methods for microscopic traffic simulation. In: Proceedings of WCTR, World Confer ence on Transport Research (WCTR) Istanbul, Turkey. Hidas, P. (2005). Modeling vehicle interactions in microscopic traffic simulation of merging and weaving. Transportation Research 13C pp. 37-62. Honeywell Inc. (2005). HTRD 400 System Reference Guide. Hoogendoorn, S.P., Zuylen, H.J.V., Schreuder, M., Gorte, B., and Vosselman, G. (2003). Microscopic Traffic Data Collection by Remote Sensing. Transportation Research Record 1855, 121-128. Hosmer, D.W. and Lemeshow, S. (2000). Applied logistic regression (2nd ed.). John Wiley and Sons Inc ., Hoboken, NJ. Hunt, J.G., and Lyons, G.D. (1994). Modeling du al carriageway lane changing using neural network. Transportation Research 2C pp. 231. ITT Industries, Inc., Systems Division (2003). T SIS version 5.1 users guide volumes 1-3. Federal Highway Administrati on Contract No. DTFH61-01-C-00005. Jha, M., Cuneo, D., and Ben-Akiva, M. (1999). Evaluation of freeway lane control for incident m anagement. Journal of Transportation Engineering ASCE 125(6), pp. 495-501. Jourard, S.M. (1964). The transpar ent self. New York: Van Nostrand. Kanungo, T., Mount, D.M., Netanyahu, N. S., Piat ko, C. D., Silverman, R., and Wu, A. Y. (2002). An efficient k-means clustering al gorithm: Analysis and implementation. IEEE Transactions on Pattern Analysis and Machine Intelligence 24 (7), pp. 881-892.

PAGE 232

232 Kesting, A., Treiber, M., and Helbing, D. (2007). General lane-chang ing model MOBIL for carfollowing models. Transportation Research Record 1999 pp. 86-94. Khattak, A., Polydoropoulou, A., and Ben-Akiva, M. (1996). Modeling reve aled and Stated pretrip travel response to advanced traveler information systems. Transportation Research Record 1537, pp. 46-54. Kinnear, T.C., Taylor, J.R., and Ahmed, S.A. (1974). Ecological ly concerned consumers: Who are they? The Journal of Marketing 38(2), pp. 20-24. Kita, H. (1999). A merging-giveway interactio n model of cars in a merging section: A game theoretic analysis. Transportation Research 33A pp.305-312. Kita, H., Tanimoto, K., and Fukuyama, K. (2002) A game theoretic analysis of merginggiveway interaction: A joint estimati on model. In: Taylor, M.A.P. (Ed.), Transportation and Traffic Theory in the 21st Century Pergamon, Oxford, pp. 503-518. Krueger, R.A., and Casey, M.A. (2000) Focus groups: A practical guide for applied research (Ed. 3rd). Sage Publications, Inc., Thousand Oaks, CA. Laval, J.A., and Daganzo, C.F. (2006) Lane-changing in traffic streams. Transportation Research 40B pp. 251-264. Levinson, D., Harder, K., Bloomfield, J, and Winiarczyk, K. (2004). Weighting waiting: evaluating perception of in-v ehicle travel tim e under moving and stopped condtions. Transportation Research Record 1898 pp. 61-68. Liu, R., Vliet, D.V., and Watling, D. (2006). Micro-simulation models incorporating both demand and supply dynamics. Transportation Research 40A pp. 125-150. Loukopoulos, P., Jakobsson, C., Garling, T., Schne ider, C.M., and Fujii, S. (2004). Car-user responses to travel demand management meas ures: Goal setting and choice of adaptation alternatives. Transportation Research 9D pp. 263. Loukopoulos, P. (2005). Future urban sustainable mobility: Implementing and understanding the impacts of policies designed to reduce priv ate automobile usage. Ph.D. Dissertation, Department of Psychology, Goteborg University, Sweden. Mahmassani, H., and Sheffi, Y. (1981) Using gap sequences to estimate gap acceptance functions. Transportation Research 15B pp. 143-148. Morgan, D.L. (1997) The Focus Group Guidebook (Focus Group Kit) Ed. 1st. Sage Publications, Inc., Thousand Oaks, CA.

PAGE 233

233 Nakanishi, M., and Cooper, L.G. (1974) Param eter estimation for a multiplicative competitive interaction model: Leas t squares approach. Journal of Marketing Research 11(3), pp. 303311. Newell, G. (1961). Nonlinear effects in the dynamics of car following. Operations Research 9, pp. 209-229. Norusis, M. (2005). SPSS 13.0 guide to data analysis. Prentice Hall. Park, B., and Qi, H. (2005). Development and ev aluation of a procedure for the calibration of sim ulation models. Transportation Research Record 1934 pp. 208-217. Peot, M.A., and Smith, D.E. (1992). Conditional nonlinear planning. P roceedings of the First International Conference on Arti ficial Intelligence Planning Systems pp. 189-197. Pet-Armacost, J.J., Sepulveda, J., and Sakude, M. (1999). Monte Carlo sensitivity analysis of unknown parameters in hazardous material s transportation risk assessment. Risk Analysis 19(6), pp. 1173-1184. Pipes, L.A. (1953). An operational analysis of traffic dynamics. Journal of Applied Physics 24 pp. 274-281. Pomerleau, D.A. (1992). Progr ess in neural network-base d vision for autonomous robot driving. In: Proceedings of the 1992 Intelligent Vehicle Symposium Detroit, MI, pp. 391396. Prassas, E S. (2002). Comparison of delay on arterials by simulation a nd HCM com putations. Transportation Research Record 1802 pp. 69-76. Prevedouros, P.D. and Wang, Y. (1999). Simulati on of large freeway and arterial network with CORSIM, INTEGRATION and WATSim. Transportation Research Record 1678, pp. 197207. Psarianos, B., Paradisis, D., Nakos, B., and Karra s, G. (2001). A cost-effective road surveying method for the assessment of road alignments. In: Proceedings of the IV International Symposium Turkish-German Joint Geodetic Days Berlin, April 3-6. PTV. (2004). VISSIM 4.00 user manual. Karlsluhe, Germany. Quadstone. (2004). Paramics V5.0, technical notes. Version No. 1.0. Rakha, H., and Zhang, Y. (2005). The INTEG RATION 2.30 framework for modeling lanechanging behavior in weaving sections. Transportation Research Record 1883 pp. 140-149. Ramanujam, V. (2007). Lane changing models for ar terial traffic. M.S. Thesis, Departm ent of Civil and Environmental Engineering, MIT.

PAGE 234

234 Ramanujam, V., Choudhury, C. F., and Ben-Akiva, M. E (2008). Extended lane-changing model to capture lane-change duration. Transportation R esearch Board 84th Annual Meeting Washington, DC. Roess, R.P., Prassas, E.S., and McShane, W.R. (2004). Tra ffic engineering (ed. 3rd). PrenticeHall. Rorbech, J. (1976). Multilane traffic flow pro cess: Evaluation of queuing and lane changing patterns. Transportation Research Record 596 pp. 22-29. Salvucci, D.D. and Liu, A. (2002). The time c ourse of a lane change: Driver control eyemovement behavior. Transportation Research 5F pp.123-132. Sheffi, Y. (1985). Urban transportation networ ks equilibrium analysis with mathematical programming methods. Prentice-Hall. Simon, H.A. (1955). A behavior al model of rational choice. Quarterly Journal of Economics, 69 (1), pp. 99-118. Simon, H.A. (1996). The scien ces of the artificial (ed. 3rd). The MIT Press, Cambridge, MA. Stevens, W. R. (1990). UNIX ne twork programming. Prentice Hall. Stevens, W. R. (1998). UNIX network programming, volume 1 (ed. 2nd): Networking APIs: sockets and XTI. Prentice Hall, 1998. Sun, J.D., Kondyli, A. and Elefteriadou, L. ( 2008). Modeling vehicle inte ractions during lanechanging maneuvers under congested conditions on arterial streets. 10th International Conference on Applications of Advan ced Technologies in Transportation (AATT2008), May 27, Athens, Greece. Sun, D.J., and Elefteriadou, L. (2009). Lane ch anging behavior on urban streets: A focus groupbased study. submit to Journal of Transportation Engineering ASCE Stewart, D.W, Shamdasani, P.N. and Rook, D. W. (2007). Focus groups: Theory and practice (2nd ed.). Sage Publications, Inc., Thousand Oaks, CA. Tang, T., Huang, H., Wang, S.C., and Jiang, R. (2007) Lane changing analysis for two-lane traffic flow. Acta Mechanica Sinica 23(1), pp 49-54. Thiemann, C., Treiber, M., and Kesting A. ( 2008) Estimating acceleration and lane-changing dynamics based on NGSIM trajectory data. Transportation Research Record 2088 pp. 90101. Tibshirani, R., Walther, G., and Ha stie, T. (2001). Estimating the number of clusters in a data set via the gap statistic. Journal of the Royal Statistical Society B 63(2), pp. 30-38.

PAGE 235

235 Toledo, T., Koutsopoulos, H.N., and Ben-Akiva, M.E. (2003). Modeling integrated lanechanging behavior. Transportation Research Record 1857 pp. 411-423. Toledo, T. (2003). Integrated driving behavior modeling. Ph.D. Thesis, Department of Civil and Environmental Engineering, MIT. Toledo, T., Ben-Akiva, M.E, Darda, D., Jha, M., and Koutsopoulos, H.N. (2004). Calibration of microscopic traffic simulation models with aggregate data. Transportation Research Record 1876, pp. 10-19. Toledo, T, and Koutsopoulos, H. N. (2004) Statistical validation of tra ffic simulation models. Transportation Research Record 1876, pp. 142-150. Toledo, T., Choudhury, C.F., and Ben-Akiva, M.E. (2005). A lane-changing model with explicit target lane choice. Transportation Research Record 1934 pp. 157-165. Toledo, T. (2007). Driving beha vior: Models and challenges. Transport Reviews 27(1), pp. 65-84. Toledo, T., and Zohar, D. (2007). Mode ling the duration of lane changes. Transportation Research Record 1999 pp. 71-78. Transportation Research Board. (1997). Highw ay Capacity Manual (HCM 1994), Washington, DC. Transportation Research Board. (2000). Highw ay Capacity Manual (HCM 2000), Washington, DC. TSS-Transport Simulation Systems. (2004). A IMSUN 5.0 Micro-simulator users manual 5.0.5. Walker, J.L. (2001). Extended discrete choice m odels: Integrated framework, flexible error sructures and latent variables. Ph.D. Thes is, Department of Civil and Environmental Engineering, MIT. Wang, J., Liu, R., and Montgomery, F. (2005). A simulation laboratory for motorway merging behavior. In: Mahmassani, H (ed.) Transporta tion and Traffic Theory: Flows, Dynamics and Human Interaction, Elsevier, 5(3), pp.127-140. Washburn, S.S. and Ko, B. (2007). Travel time re liability and truck level of service on the strategic intermodal system. Final Reports TRC-FDOT-54045-2007. Washburn, S.S. and Kondyli, A. (2006). Develop ment of guideline for driveway location and median configuration in the vicinity of interchanges. Final Reports TRC-FDOT-036-2006.

PAGE 236

236 Wei, H., Lee, J., Li, Q., and Li, C.J. (2000) Observation-based la ne-vehicle-assignment hierarchy for microscopic simulatio n on an urban street network. Transportation Research Record 1710, pp. 96-103. Wisconsin Department of Transportation, Who is an aggressive driver? http://www.dot.wisconsin.gov/safety/motoris t/behaviors/aggressive/who.htm Dec., 2006 Yang, Q., and Koutsopoulos, H.N. (1996). A micr oscopic traffic simulation for evaluation of dynamic traffic management systems. Transportation Research, 4C pp. 113129. Yang Q., Koutsopoulos H.N., and Ben-Akiva M. (2000) A simula tion laboratory for evaluating dynamic traffic management systems. Transportation Research Record 1710 pp. 122-130. Yeo, H., Skabardonis, A., Halkias, J. A., Colyar, J., and Alexiadis, V. (2008). Oversaturated freeway flow algorithm for use in next generation simulation. Transportation Research Record 2088, pp. 68-79. Zhang, Y., Owen, L.E., and Clark, J.E. (1998). Multi-regime approach for microscopic traffic simulation. Transportation Research Record 1631, pp. 103-112.

PAGE 237

237 BIOGRAPHICAL SKETCH Daniel(J ian) Sun received his Ph.D. from Transp ortation Engineering at the University of Florida in 2009. Before he started his Ph.D. study, Daniel(Jian) Sun received his bachelors and masters degrees in China in the 2000 and 2003, respectively. He has published more than 10 journal and conference papers rela ted to information integration for the Chinese railway in his masters degree study. His current major rese arch interest is in modeling the driver characteristics in microscopic simulation. He ha s already published one pape r in a related area in the10th International Conference on A pplication of Advanced Tec hnologies in Transportation, and has several papers submitted to the peer reviewed journals. Additionally, his research interest also includes urban tran sportation planning, traffic signal and traffic control. He received the Bill & Bryon Bushnell Graduate Fellowship in 2008. In his spare time, he is a fan of many sports activities. He likes to play badminton, tennis, and volleyball, and watches a lot of soccer games, as well as football and basketball games whenever GATORS are involved.