Citation
Identification, Measurement and Incorporation of Human Factors in Traffic Microsimulation Using Driving Simulator Observations

Material Information

Title:
Identification, Measurement and Incorporation of Human Factors in Traffic Microsimulation Using Driving Simulator Observations
Creator:
Manjunatha, Pruthvi
Publisher:
University of Florida
Publication Date:
Language:
English

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Civil Engineering
Civil and Coastal Engineering
Committee Chair:
ELEFTERIADOU,AGELIKI
Committee Co-Chair:
WASHBURN,SCOTT STUART
Committee Members:
SRINIVASAN,SIVARAMAKRISHNAN
STEINER,RUTH LORRAINE
Graduation Date:
8/11/2018

Subjects

Subjects / Keywords:
cognition
driverbehavior
drivingsimulator
microsimulation
traffic

Notes

General Note:
Human factors have significant influence over traffic operations. In order to understand situations such as traffic breakdowns, capacity drop etc. it is essential that human factors are measured and modeled accurately. While existing microsimulation models use stochastic distributions to represent human variability, there is lack of literature in traffic engineering that links these distributions to specific driver behaviors. At the same time, considerable work has been reported in the field of psychology on understanding, modeling and predicting human behavior and drivers intended actions. The second chapter of this dissertation is dedicated to understanding the state of the art in traditional car following, followed by a review of driver behavior models in psychology. Then, human factors to represent driver states are identified and the methods to measure them are discussed. In the third chapter a methodology is presented to use driving simulator observations in order to measure the human factor elements identified from traffic psychology. This is followed by a novel estimation methodology to address the limitations of past calibration methods discussed in the literature review. Next, the driving simulator experiment results are discussed in three parts. First, the measured driver states (Workload and Situational Awareness) is used and a driver classification is made (Chapter 5). Second, the driver behavior (Car Following, Lane Changing) is examined using trajectory data, the newly created driver classes are compared, and a modeling option is explored for a direct relationship between driver states and driver behavior (Chapter 6). Third, driver characteristics are examined using pre-screen questionnaire data and a post-experiment Driver Behavior Questionnaire (Chapter 7). The last chapter highlights the findings and presents options for future research using the framework established in this dissertation.

Record Information

Source Institution:
UFRGP
Rights Management:
All applicable rights reserved by the source institution and holding location.
Embargo Date:
8/31/2020

Downloads

This item has the following downloads:


Full Text

PAGE 1

IDENTIFICATION, MEASUREMENT AND INCORPORATION OF HUMAN FACTORS IN TRAFFIC MICROSIMULATION USING DRIVING SIMULATOR OBSERVATIONS By PRUTHVI MANJUN A THA A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2018

PAGE 2

2018 Pruthvi Manjunatha

PAGE 3

To my first teacher, my Grandmother To friends and family

PAGE 4

4 ACKNOWLEDGMENTS I thank my advisor Dr. Lily Elefteriadou, and my committee members Dr. Scott Washburn, Dr. Siva ramakrishnan Srinivasan and Dr. Ruth Steiner for their invaluable inputs. I thank Jason Rogers from Department of Public Health at UF for helping with the driving simulator. I thank Dr. Alexandra Kondyli for helping me during the initial phase of the research.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ ............... 4 LIST OF TABLES ................................ ................................ ................................ ........................... 8 LIST OF FIGURES ................................ ................................ ................................ ......................... 9 LIST OF ABBREV IATIONS ................................ ................................ ................................ ........ 11 ABSTRACT ................................ ................................ ................................ ................................ ... 12 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .................. 14 2 LITERATURE REVIEW ................................ ................................ ................................ ....... 16 Car Foll owing Models And Microsimulation ................................ ................................ ......... 16 Introduction To Car Following ................................ ................................ ........................ 16 Classical Car Following Models ................................ ................................ ..................... 17 Gazis Herman Rothery (GHR) models ................................ ................................ .... 17 Collision avoidance (CA) models ................................ ................................ ............ 17 Action point (AP) models ................................ ................................ ........................ 18 Mode rn Car Following Models ................................ ................................ ....................... 19 Optimum velocity models (OVM) ................................ ................................ ........... 20 Desired measures models ................................ ................................ ......................... 21 Trajectory based models ................................ ................................ ........................... 22 Behavioral Parameters In Microsimulation ................................ ................................ ..... 24 Driver variability ................................ ................................ ................................ ...... 24 Car following ................................ ................................ ................................ ........... 25 Gap acceptance and lane changing ................................ ................................ ........... 26 Discussion ................................ ................................ ................................ ................ 27 Driver Behavior Models In Psychology ................................ ................................ ................. 27 Lower Level Control (Maneuvering And Control) ................................ ......................... 29 Higher Level Control (Task Management, Decision Making) ................................ ........ 30 Application of TCI concept in Car Following (CF) ................................ ........................ 32 Situational Awareness (SA) And Driver Workload (WL) ................................ .............. 35 Measuring Driver Behavior Parameters ................................ ................................ ................. 37 Overview ................................ ................................ ................................ ......................... 37 Measuring Driver Workload ................................ ................................ ............................ 38 Self report measures ................................ ................................ ................................ 38 Secondary task measures ................................ ................................ .......................... 39 Physiological measures ................................ ................................ ............................ 39 Measuring Situational Awareness ................................ ................................ ................... 40 Summary Of WL And SA Measurement Studies ................................ ............................ 40

PAGE 6

6 Measuring Driver Characteristics ................................ ................................ .................... 42 Instrumented vehicles ................................ ................................ ............................... 42 Focus group study ................................ ................................ ................................ .... 42 Driver behavior questionnaire (DBQ) ................................ ................................ ...... 43 Discussion ................................ ................................ ................................ ............................... 43 3 METHODOLOGY AND DATA COLLECTION ................................ ................................ 45 Experiment Equipment ................................ ................................ ................................ ........... 48 Participants ................................ ................................ ................................ ............................. 49 Experimental Design And Driving Scenarios ................................ ................................ ......... 49 4 A NOVEL METHOD FOR ESTIMATION OF CAR FOLLOWING PARAMETERS ....... 53 Relationship Between Wiedemann Parameters And Regime Thresholds .............................. 54 Measurement Methodology ................................ ................................ ................................ .... 55 Estimation Results ................................ ................................ ................................ .................. 57 Estimation Discussion ................................ ................................ ................................ ............ 59 5 DRIVER STATES: WORKLOAD AND SITUATIONAL AWARENESS .......................... 61 RQ1 Do Types Of Freeway Sections Influence Driver States (WL And SA) And If Yes, How? ................................ ................................ ................................ ................................ ... 62 RQ2 Do Changes In Traffic Density Influence Driver States (WL And SA) And If Yes, How? ................................ ................................ ................................ ................................ ... 63 RQ3 Do Drivers Use Compensatory Mechanisms And What Is The Effect Of Such Compensation? ................................ ................................ ................................ .................... 64 RQ4 Are Different Methods Of Workload Measurement (NASA TLX And PDT) Consistent? ................................ ................................ ................................ .......................... 65 RQ5 Can Drivers Be Classified Based On WL And SA States, If So What Are The Properties Of These Classes? ................................ ................................ .............................. 68 RQ6 How Does The Relationship Between WL & SA Vary Across Freeway Conditions For Different Driver Classes? ................................ ................................ ............................. 72 Discussion On Driver States ................................ ................................ ................................ ... 74 6 DRIVER BEHAVIOR: CAR FOLLOWING AND LANE CHANGING ............................. 76 RQ1 Do Drivers Classified Based On WL And SA States, Have Different CF Behavior? ... 77 RQ2 Do Types Of Freeway Sections Influence Driver Behavior And If Yes, How? ............ 78 RQ3 Do Changes In Traffic Density Influence Driver Behavior And If Yes, How? ............. 80 RQ4 How Do Driving Regimes And The Minimum Safe Headway Vary By Driver Group And Traffic Condition? ................................ ................................ ............................ 82 RQ5 How To Model The Relationship Between Driver States And Driver Behavior? ......... 87 Model ................................ ................................ ................................ ....................... 88 Boundary Conditions And Physical Meaning ................................ .......................... 89 Model Discussion ................................ ................................ ................................ ................... 90 7 DRIVER BEHAVIOR QUESTIONNAIRE (DBQ) SURVEY ................................ ............. 92

PAGE 7

7 Driver Group Properties ................................ ................................ ................................ ......... 92 DBQ Responses ................................ ................................ ................................ ...................... 94 Factor Analysis ................................ ................................ ................................ ....................... 96 Principal Co mponents Analysis (PCA) ................................ ................................ .... 97 Factor 1 Situational Awareness/Distraction ................................ ........................... 100 Factor 2 Workload Adjustment/Compensation ................................ ...................... 101 Factor 3 Errors And Misjudgments ................................ ................................ ........ 102 Factor Analysis Discussion ................................ ................................ ................................ ... 103 8 CONCLUSIONS AND FUTURE WORK ................................ ................................ ........... 104 Summary ................................ ................................ ................................ ............................... 105 Conclusions ................................ ................................ ................................ ........................... 106 Future Work ................................ ................................ ................................ .......................... 106 APPENDIX A PARTICIPANT RECRUITMENT FLYER ................................ ................................ ......... 109 B PRESCREEN QUESTIONNAIRE ................................ ................................ ...................... 110 C SIMULATOR SICKNESS QUESTIONNAIRE (SSQ) ................................ ....................... 113 D NASA TLX QUESTIONS ................................ ................................ ................................ ... 114 E SITUATIONAL AWARENESS GLOBAL ASSESSMENT TEST (SAGAT) QUESTIONS ................................ ................................ ................................ ........................ 116 F DRIVER BEHAVIOR QUESTIONNAIRE (DBQ) ................................ ............................ 117 G PERIPHERAL DETECTION TEST (PDT) ................................ ................................ ......... 118 LIST OF REFERENCES ................................ ................................ ................................ ............. 119 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 126

PAGE 8

8 LIST OF TABLES Table page 2 1 Terminology Used in Literature ................................ ................................ ......................... 37 3 1 Traffic States ................................ ................................ ................................ ...................... 50 4 1 Wiedemann parameters at different speeds for Driver 12 ................................ ................. 57 5 1 Average weights calculated from NASA TLX self rating responses ................................ 66

PAGE 9

9 LIST OF FIGURES Figure page 2 1 Wiedemann Thresholds and Regimes ................................ ................................ ................ 19 2 2 ................................ ................................ ................................ .................. 22 2 3 Summary of Behavioral Parameters Used in the Microsimulation Packages .................... 25 2 4 The hierarchical structure of the road user task (Michon, 1985) ................................ ....... 28 2 5 Two way classification of driver behavior models (Michon, 1985) ................................ .. 29 2 6 Higher level control models ................................ ................................ ............................... 31 2 7 Relationship among factors affecting driving (Carsten, 2007) ................................ .......... 36 2 8 ................................ ...... 36 2 9 Driver Behavior Data Collection Methods ................................ ................................ ........ 38 2 10 Summary of the Studies with Driver WL and SA Measurement ................................ ....... 41 3 1 Elements of Traffic Operations ................................ ................................ .......................... 45 3 2 Driving Simulator ................................ ................................ ................................ .............. 48 3 3 Gender Age Mix in the group of participants ................................ ................................ .... 49 3 4 Measurement Sections ................................ ................................ ................................ ....... 50 3 5 Driving Scenarios on a Freeway for a given density ................................ ......................... 51 4 1 Wiedemann parameters and regimes plot used in VISSIM ................................ ............... 54 4 2 Trajectory and thresholds as per default VISSIM values v/s actual trajectory .................. 55 4 3 ................................ ............................ 56 4 4 Wiedemann parameters estimated for different speeds for all drivers .............................. 58 4 5 Wiedemann paramete rs by age group ................................ ................................ ................ 58 4 6 Wiedemann parameters by driver aggressiveness ................................ ............................. 59 5 1 Driver States by Type of Freeway Segment ................................ ................................ ...... 62 5 2 Driver States by Level of Traffic Density ................................ ................................ .......... 63

PAGE 10

10 5 3 Mean Speed as an Indicator of Compensatory Mechanism ................................ ............... 64 5 4 NASA TLX Self Rating Category Averages ................................ ................................ ..... 66 5 5 Relationship between WL (PDT) and WL (NASA TLX) measures ................................ 67 5 6 Clustering Statistics ................................ ................................ ................................ ........... 69 5 7 Mean WL (PDT) and SA for the clustered groups ................................ ............................ 69 5 8 Properties of driver classes based on pre screening questionnaire ................................ .... 70 5 9 Driver State Sub Lev els by Driver Groups ................................ ................................ ........ 71 5 1 0 Relationship between WL and SA across different conditions ................................ ......... 73 6 1 Car Following Parameters (Estimated for Control Scenario) by Driver Group ................ 77 6 2 Longitudinal Behavior by Freeway Segment ................................ ................................ .... 79 6 3 Lateral Behavior by Freeway Segment ................................ ................................ .............. 79 6 4 Longitudinal Behavior by Traffic Density ................................ ................................ ......... 81 6 5 Lateral Behavior by Traffic Density ................................ ................................ .................. 82 6 6 Driving Regimes by Density ................................ ................................ .............................. 84 6 7 Minimum Safe Headway by Level of Traffic and Type of Segment ................................ 86 6 8 Relationship between CC1 and (WL/SA) ................................ ................................ .......... 88 7 1 Properties of driver classes based on pre screening questionnaire ................................ .... 93 7 2 DBQ Responses by Driver Groups ................................ ................................ .................... 95 7 3 Eigenvalue Scree Plot for DBQ data ................................ ................................ ................. 98 7 4 Total Variance Explained by Each of the Components ................................ ..................... 99 7 5 Rotated Component Matrix with Factor Loadings ................................ ............................ 99 7 6 Factor 1 Responses ................................ ................................ ................................ .......... 100 7 7 Factor 2 Responses ................................ ................................ ................................ .......... 101 7 8 Factor 3 Responses ................................ ................................ ................................ .......... 102 8 1 Elements of Traffic Operations (Reproduced from 3 1) ................................ ................. 104

PAGE 11

11 LIST OF ABBREVIATIONS DBQ Driver Behavior Questionnaire DC Driver Characteristic PDT Peripheral Detection Task SA Situational Awareness SSQ Simulator Sickness Questionnaire TCI Task Capability Interface WL Driver Workload

PAGE 12

12 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy IDENTIFICATION, MEASUREMENT AND INCORPORATION OF HUMAN FACTORS IN TRAFFIC MICROSIMULATION USING DRIVING SIMULATOR OBSERVATIONS By Pruthvi Manjunatha August 2018 Chair: Lily Elefteriadou Major: Civil Engineering Human factors have significant influence over traffic operations. In order to understand situations such as traffic breakdowns, capacity drop etc. it is essential that human factors are measured and modeled accurately. While existing microsimulation models use stochastic distributions to represent human variability there is lack of literature in traffic engineering that links these distributions to specific driver behaviors. At the same time, considerable work has been reported in the field of psychology on understanding, modeling and predicting human The second chapter of this dissertation is dedicated to understanding the state of the art in traditional car following, followed by a review of driver behavior models in psychology. Then, h uman factors to represent driver states are identified and the methods to me asure them are discussed. In the third chapter a methodology is presented to use driving simulator observations in order to measure the human factor elements identified from traffic psychology This is followed by a novel estimation methodology to address the limitations of past calibration methods discussed in the literature review. Next the driving simulator experiment results are discussed in three parts. First the measured driver states (Workload and Situational Awareness) is used and a driver classification

PAGE 13

13 is made (Chapter 5) Second, the driver behavior (Car Following, Lane Changing) is examined using trajectory data the newly created driver classes are compared and a modeling option is explored for a direct relationship between driver states and driver behavior (Chapter 6) Third, driver c haracteristics are examined using pre screen questionnaire data and a post experiment Driver Behavior Questionnaire (Chapter 7) The last chapter highlight s the finding s and present s options for future research using the framework established in this dissertation.

PAGE 14

14 CHAPTER 1 INTRODUCTION The earliest car following models were developed by traffic engineers to study the effects of traffic flow as a whole. Though these models have greater mathematical fidelity with traffic flow equations, they are unable to replicate certain traffic phenomena like capacity drop, breakdown etc. This can be directly attributed to l ack of representation of human factors (Saifuzzaman and Zheng, 201 4) Since the initial theorization and development in 1950s, car following models have evolved and have been co nstantly modified by traffic engineers to explain various traffic phenomenon. However, one common issue with many of these models is that resemble general function approximators or ad hoc control laws than psychologically plausible char acterizations of how humans think about and solve the driving problem (Boer, 1999 ). Secondly, the parameters in the traffic engineering models that are supposed to represent behavior are calibrated based on minimizing errors between observed and simulated aggregate traffic measures such as averag e speed, travel time and delay Hence they do not represent the intended behavior (Manjunatha et al, 2013) Thus, the objective of this dissertation is to make a clear conne ction between these behavior parameters and the behavior they are supposed to represent. Based on the literature review, t here is a need to develop methods which directly address and replicate the behavior associated with the concerned parameters. At the same time traffic psychologists, occupational therapists, and traffic safety researchers have been studying and developing their own models of driving behavior (Manjunatha et al, 2017). These models are not focused on traffic operational characteristics. Rather, they have been developed to understand the psychology of driving under various

PAGE 15

15 conditions such as distracted driving, adaptive cruise control etc (Ma and Kaber, 2005). Though these psychological models use different approaches and terminologies, there are commons concepts such as Driver Workload (WL), Situational Awareness (SA) etc. that have the potential of bridging the gap between models in psychology and traffic engineering. The se psychological concep ts and their inter relationships need to be studied and modeled. There have been some studies recently (Hoogendoorn et al 2013 Saifuzzaman et al 2015 ) that have presented theoretical frameworks to incorporate some of the concepts from psychology in traffic engineering models. However, these frameworks hav e not been validated with field data. There are a number of methods in psychology (questionnaires, surveys, driving simulator and instrumented vehicle measurements etc.) to collect driver behavior data (Manjunatha et al, 2017) Using controlled environment s in a driving simulator behavior data can be directly measured fr o m the drivers. Hence, the final objective of the dissertation is to develop behavior based car following models using data collected directly from the drivers that would be consistent with the driver behavior models in psychology as well as models in traffic engineering. In the following chapter, a literature review is conducted to understand the sta te of the art of the driver behavior in traffic engineering and psychological models followed by methodology, data collection and results.

PAGE 16

16 CHAPTER 2 LITERATURE REVIEW The literature review section has been divided in to three parts T he First part reviews the evolution of car following models and their implementation in traffic microsimulation packages is studied. The advantages and limitation of models in traffic engineering are discussed. T he second part reviews driver behavior models in psychology the gaps between psychology and traffic engineering are studied and the potential ways to bridge those gaps are identified. T he third part presents various methods to collect driver behavior data with advantages and limitatio n s of each method C ar Following Models And Microsimulation Introduction To Car F ollowing T raffic microsimulation models simulate the behavior of individual vehicles within a predefined road network and are used for the evaluation and development of road traffic management and control Car following (CF) models are fundamental components of traf fic microsimulation. CF models define the longitudinal interactions between the vehicles on the road. The concept of car following was first introduced by Pipes (1953), where it was assumed that the distance of the following vehicle to the lead vehicle is equal to one vehicle length for every ten miles per hour of the subject vehicle speed. Since then, the car following models have evolved based on different approaches. The car following approaches developed from 1950s to early 1990s have been reviewed unde r the section Classical Car following models, whereas the more recent approaches are discussed under Modern Car following models.

PAGE 17

17 Classical Car Following Models The classical car following models are grouped into three classes based on the principles of the original model they were improved upon. Each of them are discussed in the following sub sections. Gazis Herman Rothery (GHR) models These are single regime stimulus acceleration to relative speed and rela tive distances (Gazis et al 1961). These models considered the sensitivity of the reaction to the lead vehicle, and they were of the form: ( 2 1) is the next time interval (it can also be interpreted as the reaction time), m,l represented a sensitivity parameter, and m and l are calibration exponents. While simplicity of this model i s an advantage, it has several limitations including inability to adequately model inter driver heterogeneity. Different parameters were added and improvements were built upon this model. MITSIM model (Yang and Koutsopoulos, 1996), a class of GHR models, e xplicitly considers multiple regimes in car following. Collision avoidance (CA) models f the most widely used models built on this principle (Gipps, 1981). Presently, AIMSUN traffic simulator (TSS, 2005) uses this model. The model calculates two speeds: the speed of the following vehicle under non constrained conditions, and the s peed that would result if the following vehicle is constrained by

PAGE 18

18 the lead vehicle. The minimum of these two speeds is selected as the follower vehicle speed. min (2 2) where, is the speed of the vehicle n+1 at the time t+ is the apparent reaction time, a constant for all vehicles is the speed of the vehicle n+1 at time t is the maximum acceleration which the driver of vehicle n+1 wishes to undertake is the speed at which the driver of vehicle n+1 wishes to travel is the actual most severe deceleration rate that the driver of vehicle n+1 wishes to undertake location of the front vehicle n at t location of the subject vehicle at time t L(n) is the effective size of the vehicle n b n is the most severe deceleration Speed of subject vehicle at time t. Action point (AP) models Also called psycho physical models, these models define different modes of driving based on some perception thresholds. Presently, VISSIM traffic simulator uses W i e model (Wiedemann, 1974) (Fritzsche, 1994 ) In th e Wiedemann model acceleration function changes depending on the driving regime. The so called perception thresholds define fou r different driving regimes: ( 1 ) free flow, ( 2 )

PAGE 19

19 approaching, ( 3 ) following ( 4 ) decelerating. The car following parameters in the microsimulation package VISSIM define these thresholds (Figure 2 1). Figure 2 1 Wiedemann Thresholds and Regimes Brackstone and McDonald (1999) were one of the earliest to comprehensively review the classical car following models and describe their limitations in modeling human behavior. There have been many other such reviews since then (Panwai and Dia, 2005; Hamda r, 2012) often including classical and some modern models. Different models have different limitation s based on which However, one common issue with many of these models is that they look more like general function approximators or ad hoc control laws than psychologically plausible characterizations of how humans think about and solve the driving problem (Boer, 2000). From late 1990s till now, there have been many efforts to address the limitat ions in the classical car following models. The following sub section discusses those models in brief. Modern Car Following Models The modern car following models have been grouped into three classes named Optimum Velocity, Desired Measures and Trajectory Based Models respectively. Each of them are discussed in the following sub sections.

PAGE 20

20 Optimum v elocity m odels (OVM) Bando et al. (1995) stated that each driver tries to achieve an optimal velocity based on the distance to the preceding vehicle and the speed difference between the vehicles. This was an alternative possibility explored recently in car following models. The formulation is based on the assumption that the desired speed depends on the distance headway of the nth vehicle (i.e the distance between the vehicle under consideration to the front vehicle) The function representing the relation between desired velocity and the instantaneous distance headway was called legal or optimal velocity function. V desired = V opt x n ( t )} ( 2 3 ) The acceleration is up dated as a n ( t ) = k [ V x n ( t )) v n ( t )] (2 4 ) where k is the sensitivity factor. The optimal velocity function has to be: 1 Monotonically increasing function with x n ( t ) 2 It has an upper bound V opt max = V opt xn ( t ) } 3 It has a turning point x n ( t ) = h c : ( h c ) = 0, where h c is the safe distance A general form optimum velocity function is given by: V x n ( t ))=0 5 v max x n ( t ) h c )+tanh( h c )] ( 2 5 ) This model was later revised by Bando et al (1998) : a n ( t + ) = k [ V x n ( t )) v n ( t )] ( 2 6 ) Ap plying the optimal velocity model (OVM), many properties described, such as the instability formation of stop and go waves But Treiber et al ( 2000 ) pointed out that if the preceding cars are much faster, then the vehicle will not brake, even if its headway is smaller than the safe distance, and this instance cannot be e xplained by optimal ve locity model. This lead to the development of new class o f models based on optimal velocity model.

PAGE 21

21 Jiang et.al (2002 ) Zh a o and Gao (2005) developed a class of models that take both positive and negative velocity differences into account and called them full velocity difference model (FVD) and full velocity and acceleration difference model (FVAD) respectively. Lenz et ptimal velocity model to consider multiple leader of the subject vehicle The vehicle directly ahead and closest to the subject vehicle is weighted the most. A detailed study was made by Lenz et al. (1999) on linear stability with hysteresis loops, flux density relations an d the speed of the fronts of a congested flow with simulation of a sample data. Due to the effect of anticipating multiple vehicles ahead, the modified model showed better stability Desired m easures m odels Intelligent driver model (IDM) (Treiber et al, 2000) is a typical modern car following model based on the desired measures. T his model considers both the desired speed and the desired space headway It assumes that the ntinuous function of the ( v n ), the spacing ( d n ) from the leader, and the speed difference from the leading v n

PAGE 22

22 This model has been modified and used by researchers to introduce specific human factors (Hoogendoorn et al, 2013 Saifuzzaman et al 2015 ). Those efforts are discussed in detail in the next section. Trajectory b ased m odels Fig ure 2 2 (a) Piecewise linear vehicle trajectories (adopted from Newell, 2002), (b) Relation between velocity and spacing for an individual driver (adopted from Newell, 2002),

PAGE 23

23 Newell (2002), stated that the following vehicle traces the same trajectory as that of the leader, except for a translation in time and space. Consider the vehicles j and j 1 in figure 1. Vehicle j follows the modified linear trajectory of j 1 with a translation of d j in space and j in time. x j j ) = x j 1 (t)+ d j ( 2 9 ) d j and j vary from driver to driver and j is the time needed for driver j to reach the preferred spacing for a new velocity The driver aggressiveness can be incorporated by the calibration of d j and j in the model. This is a very simple model with very few pa rameters to be calibrated. The F igure 2 2 (c) demonstrates that this model confirms with the Lighthill Whitham Richards (LWR 1955 ) Model However in this form, it is applicable only in homogenous conditions with very few lane trajectories. Also with respect to stability, the waves neither amplify nor decay, inste ad simply propagate. Ahn et al (2004) verified this model with a set of field data. As expected, for a case with geometric inhomogeneity, the model did not perform well. There have been a number of st udies comparing variants of both modern and classical ca r following models. Punzo and Simonelli (2005) compared intelligent driver model (IDM) and Italy. Headway, speed and spacing were used as measures of performance (MOPs) with RMSPe ry good performance in homogenous conditions. Ranjitkar et al. (2005) conducted comparison of eight car following models with a

PAGE 24

24 test track: Chandler Generalized General Motors ( GGM ), Gipps Krauss OVM Newell CA Leutzbach Chandler and GGM models perfo rmed well overall, but t he modified optimum velocity model (OVM) performed well with speed and acceleration predictions The Newell model produced competitive percentile error values for spacing prediction, while the same for speed and acceleration predict ion were relatively higher. Since different models have their own advantages and limitations, model performances can vary significantly based on experimental setup, measures of performance chosen and type of statistical test conducted. Behavioral Parameters In M icrosimulation While the previous section reviewed the evolution of car following models, t his section summarizes the consideration of driver behavior in microsimulation packages. Hidas ( 2005 ) highlighted the difference between three simul ation packages (VISSIM, AIMSUN and PARAMICS) based on how individual drivers are modeled. Figure 2 3 presents an update of that work for recent versions of these packages along with CORSIM and MITSIM. This list is not exhaustive; the intent here is to exam ine several representative packages and evaluate the opportunities available for enhancements in traffic modeling. Driver v ariability Driver behavior and variability in AIMSUN (TSS, 2009), VISSIM (PTV, 2014) and PARAMICS (Quadstone, 2004) is modeled join tly with vehicle types (Barcelo, 2010). The vehicle attributes and driver attributes are considered jointly in Driver Vehicle Units (DVU). It is possible to model driver behavior in these packages by considering each vehicle type. In CORSIM (TSIS, 2011) d river type and vehicle type are modeled independently. There are 10 driver types with varying aggressiveness. In MITSIM driver variability is mainly modeled through the desired speed distribution (Ben Akiwa et al, 2002).

PAGE 25

25 Figure 2 3 Summary of Behavioral Parameters Used in the Microsimulation Packages Waiting time affects the gap acceptance in AIMSUN, which is one of i ts distinguis hing driver behavior aspects In PARAMICS three behavior specific factors are considered: awareness, aggressiveness, and patience. In VISSIM, the specific behavioral characteristics of drivers/vehicles are the psycho physical sensitivity thre sholds. Driving behavior parameter sets can be defined in VISSIM, and road links in the same network can be assigned to different driving behavior sets. Car f ollowing The car following models used in the packages are listed in Figure 2 3 Except AI MSUN and CORSIM, the others are regime based models. In PARAMICS driver aggressiveness and awareness when approaching a lane drop, as well as reaction time affect car following. VISSIM

PAGE 26

26 provides a set of ten parameters (CC0 to CC9) to define speed and dista nce thresholds between different regimes. The microsimulation packages described in Figure 2 3 mostly use variants of classical car following models and the implantation of modern car following models are generally limited to custom built environments. This is because, most of the popular microsimulation packages have However some of them (e.g: VISSIM) following through their Appli cation Programming Interface (API). Gap a cceptance and l ane c hanging Gap acceptance models give way behavior of a lower priority vehicle approaching a conflict point. AIMSUN uses maximum give way time which is a fixed constant, whereas PARAMICS uses an VISSIM the give way parameters are defined at each conflict point, i.e., there is spatial variation, but these parameters are the same for all drivers at a given conflict point. In CORSIM, the lane changing logic is different for freeway and arterial segments. Distribution of the lane changing distance across the ten driver types makes aggressive drivers more likely to make discretionary lane changes Yang and Koutsopoulos (1996 ) implemente d a rule based lane changing model into MITSIM. They presented a merging model, separately from the lane changing model. The merging model incorporates courtesy yielding. Critical review of lane changing models in CORSIM, MITSIM and other re cent studies ca n be found in Moridpour et al (2013 ).

PAGE 27

27 Discussion Driver variability is considered in all five of the packages in different ways. Though there are specific parameters within car following and lane changing for driver variability, most of these packages use statistical distributions to determine parameter values of car following and lane changing models for individual vehicles. Depending on how the distributions in microsimulation are calibrated, one can achieve different degrees of variabil ity. For example in CORSIM the ten driver types have different desired speeds, while in other packages there is only one desired speed parameter which is normally distributed. In both cases similar variability can be achieved. Even if the driver type proportion is non u niform in CORSIM, an equivalent skewed curve can be used to fit this and achieve the required variability. In Figure 2 3 the alphanumeric symbols in parentheses indicate whether the parameter has a constant value or a distribution within each driver type, and whether it is global, link specific or vehicle specific. These elements can affect whether the simulator can consider additional driver related parameters. For commercial microsimulation packages default parameters (e.g. car following variables) and their distribution (e.g. acceleration functions) are calibrated after a thorough data collection in the field ( Soria et al, 2014 ). These parameters represent unknown factors, often attributed to driver behavior elements, vehicle type, traffic state, weathe r, etc. ( Manjunatha et al, 2013 ) and their interactions. Developing relationships between these parameters and traffic operations could improve the calibration process. Driver B ehavior M odels I n P sychology A number of studies have been conducted in the field of human factors and psychology to understand driver behavior. Michon ( 1985 ) brought the attention of the traffic simulation community to the gap between traffic modeling and driver behavior. He classified various types

PAGE 28

28 of models and suggested alternative modeling approaches. Driving was explained as a hierarchical process with knowledge based (strategic), rule based (maneuvering) and skill based (control) levels (Figure 2 4 ). Figure 2 4 The hierarchical structure of the road user task (Michon, 1985) The strategic level involves the determination of trip goals, route choice, etc. The maneuvering level involves obstacle avoidance, gap acceptance, turning, overtaking, etc. In the control level, except emergencies such as braking and in case of novice drivers, the skills are performed without conscious control and use of attention resources. Control behavior is immediate and efficient, whereas maneuvering behavior occurs under conscious control and requires attentio n. As shown in Figure 2 4 an important component in both the maneuvering and control levels is the environmental input, which can be interpreted as prevailing network, traffic and weather conditions. Michon ( 1985 ) summarized various driver behavior models as shown in F igure 2 5 and suggested two ways of classification. First, input output models are behavior oriented and consider driving as an intermittent task, whereas internal state models are related to traits and

PAGE 29

29 motivational levels of the driver. Then, models with dynamic interaction were grouped as functional and the rest as taxonomic. Figure 2 5 Two way classification of driver behavior models (Michon, 1985) While highlighting the weakness of internal state models for their subjectivity and lack of field observed data, Michon (1985) recommended that a general driver behavior model may be built that takes advantage of the benefits of task models together with t he adaptive control (productions) based on declarative information accep to 50,000 of such functions may sufficiently incorporate the three levels of road performance task. In the next subsections driver behavior models are divided into lower and higher level controls according to Figure 1 and are studied to identify elements and processes that can adapted be in the car following models and microsimulation. Lower L evel C ontrol ( M aneuvering And C ontrol) Salvucci (2 006 ) used a cognitive architecture called Atomic Components of Thought Rationale (ACT

PAGE 30

30 combining the benefits of task models and the adaptive con trol dynamics. The following 1. Acquire visual info from visual processor. 2. Compute and update steering angle and acceleration based on visual angles ( and time headways ( THW ). 3. Send updates to motor system. Step 2 looks similar to car following models, but unlike car following models where the vehicle states are readily available, in this model every individual sub step (e.g., looking at front vehicle, assessing headways, making lateral and longitudinal decisions) is modeled as distra ction, driver heterogeneity, etc. but without calibration and testing. This cognitive framework requires very detailed modeling compared to microsimulation. Adapting this model to a typical traffic microsimulation package as is would require exorbitant amo unts of computational power. However, the principles of task analysis are useful. Higher L evel C ontrol ( T ask M anagement, D ecision M aking) The higher level of control involves concepts such as workload and task demand. These are similar in the sense that they relate to the quantification of the effects of the level of difficulty associated with driving, given the state and attributes of the driver and environment. To differentiate changes in driving behavior, Bekiaris et al. ( 2003 ) introduced the con cept 6 a.) The use of this concept was shown through its application in different examples.

PAGE 31

31 Figure 2 6 Higher level control models (a) Factors affecting drivability (Bekiaris, 2003), (b) Task capability interface (TCI) model (Fuller, 2005 )

PAGE 32

32 Driva bility index was defined as weighted product of factor indices. Since these factors to explore relationships among factors affecting driving behavior. Fuller ( 2005 ) introduced the task capability interface (TCI) model (Figure 2 6 b) to study the effects of task demand on risk taking. In this model, driving is considered a result of capability and task demands. The task difficulty is inversely proportional to the diff erence between task demand and driver capability. The factors affecting Capability (C) and Task Demand (D) are shown in Figure ( 2 6 b ) The task capability interface model concentrates on the human factors aspect of the driver behavior modeling. Task Dema nd depends on Capability because the driver adjusts speed, time headway and awareness whenever Task Demand Hoogendoorn et al 2013 ). There have been two no applications of the T CI model in car following ( Hoogendoorn et al, 2013 and Saifuzzaman et al, 2015 ), discussed in following subsection Application of TCI concept in Car Following (CF) ility is pitted against task demand. Hoogendorn et al ( 2013 ) and Saifuzzaman et al ( 2015 ) have applied this concept in two different ways. Hoogendoorn et al. ( 2013 ) made two attempts to incorporate task capability in car following models. They differentiated between the compensation and performance effects and incorporated them in the intelligent driver model (IDM):

PAGE 33

33 Compensation effects, m c ( t ) entail conscious adaptation effects in driving behavior in order to increase or reduce the difficulty of the driving task. This occurs when the drivers realize that workload is approaching their capability. Compensation effects (Eq n 2 14 ) in longitudinal driving behavior may consist of speed reductions and changes in the distance to the lead vehicle. Performance effects, m p ( t ) entail an increase or reduction in the performance of the driving task. This happens when drivers are multi tasking or encountering certain situations (for example, significant rearview mirror checking, braking in reaction to decelerations of the lead vehicle).

PAGE 34

34 For an assumed range of values of m c ( t ) and m p ( t ), Hoogendorn et al. ( 2013 ) simulated traffic flow along a freeway and showed the difference in driving performance between the case of optimal information and overload of information to the driver. In their second attempt, Hoogendoorn and van Arem ( 2013 ) included task difficulty as one of the inputs and made an attempt to model it using a neuro fuzzy approach based on data from a driving simulator. The range and magnitude of task demand or task difficulty were assumed. To explain the adaptation effect Hoogendoorn et al. ( 2013 SA. The interaction between task demand and driver capability in TCI was assumed to be a direct ratio called Task Difficulty (TD) by Saifuzzaman et al ( 2015) They claimed that task demand at any instance could be explained by the speed of the driven vehicle and the spacing to the preceding vehicle. They also concluded that, there is a direct correlation between driver capability and time headway selection Considering these explanations, the task difficulty w as given by:

PAGE 35

35 Situational A wareness ( SA ) And D river W orkload ( WL ) Endsley et al. ( 1995 ) defined situational awareness as the degree to which a driver is aware of the surrounding environment, and specifically those elements that are relevant for the driving task. Situational awareness can also be considered a state variable that may dynamically evolve over time as a function of Workload, Task Capability and many other factors ( Van Lint et al 2016 ). Figure 2 7 provides a schematic of the concept ac cording to Carsten ( 2007 ), who classified causal factors into long medium and short term. Long term factors (experience, attitudes) are responsible for the driver state (a medium term factor), which in turn i nfluences the term factors like workload and situational awareness. Medium and short term driver characteristics (Figure 2 7 ) correspond to actions at the strategic and maneuvering levels respectively (Figure 2 4 ).

PAGE 36

36 Literature h as consistently suggested that situational awareness and workload constantly interact with one another (Endsley 1995, Carsten, 2007, Hoogendoorn, 2013 etc). These interactions contribute to the final driving performance. F igure 2 7 Relationship among factors affecting driving ( Carsten, 2007 ) Based on 2 8 ) Through this model, de Waard (1996) broke down the inter relationship between demand, workload and performance into six different inter action zones. Figure 2 8 (de Waard, 1996)

PAGE 37

37 Figure 2 eory proposed by Hoogendoorn (2013) is similar to this model. Since studies in the literature use different terminologies for similar concepts, we have developed 2 1 to summarize these and the terminology used network, traffic and environment is called task demand their ability to comprehend the movem 2 1 (WL, SA and DC) are all measurable entities. Table 2 1 Terminology Used in Literature Net effect of Network, Traffic &Environment Awareness of the Surrounding Traffic Ability and Nature of the Driver Bekiaris et al (2003) Workload Risk Awareness Drivability Fuller (2005) Task Demand Human Factors Capability Carsten (2007) Workload & Task Demand Situation Awareness Driver State Hoogendoorn (2013) Task Demand Activation Level Driver Capability Current Study Workload (WL) Situational Awareness (SA) Driver Characteristics (DC) Measuring D river B ehavior P arameters Overview Measuring behavior is an important step in bringing the models and elements discussed in the previous sections from theory to practice. Researchers may use external observations, such as vehicle trajectories, to approximate driver behavior; however, this method by itself does not earchers have used other methods to obtain quantitative and qualitative driver behavior information, including

PAGE 38

38 instrumented vehicles, focus groups, questionnaires, and driving simulators. A detailed review of traditional data collection methods associated with microsimulation can be found in ( Auberlet et al, 2014 ). This section focuses on measuring the behavioral elements identified earlier, i.e. WL, SA and DC ( 2 1 ). Figure 2 9 shows different methods of collecting behavioral data. Driver characteristics can be collected in various methods like questionnaires, focus groups and instrumented vehicles. However, due to the obtrusive nature of the measurement methods, it is safer to collect workload and situational awareness data through driving simulator experiments. Figure 2 9 Driver Behavior Data Collection Methods Measuring D river W orkload Driver workl oad can be defined as the net effect of network, traffic and environmental conditions experienced while driving. Advantages and limitations of WL data collection methods with respect to specific measures are discussed below. Self r eport m easures Hart et al. ( 1988 ) developed a subjective measure (NASA TLX) for mental workload and used a questionnaire to assess it. In this method, six different criteria are used (Mental Demand,

PAGE 39

39 Physical Demand, Temporal Demand, Effort, Performance and Frustration). At t he end of each driving scenario, the participants are asked to rate each of the criteria individually. Later, the criteria are compared in pairs, establishing weights and the workload is calculated as a weighted mean of these ratings. The detailed document for the NASA TLX measure is attached in the appendix. Generally, questionnaires seem to be useful in identifying trends, particularly for studies requiring larger samples. But self reported measures could contain bias and may have to be cross validated o ther methods of measurement. Secondary t ask m easures Driving simulators have been extensively used in driver behavior research. Performance measures are collected by asking subjects to perform a secondary task while driving in the simulator. Experiments usually involve responding to a random audio or a visual clue ( de Waard, 1996 ). The visual cues are on either side of the screen i.e. at the periphery of the eye. Hence, it is called Peripheral Detection Task (PDT). The higher the hit rate or the correc t responses to the PDT, lower the workload and vice versa. Physiological m easures Researchers ( Chang et al, 2001, Jahn et al, 2005 ) used physiological measurements (heart rate/variability, electromyogram etc.) and visual methods (eye tracking, visual occlusion, etc.) to objectively measure mental workload. Chang et al. ( 2001 ) mentioned skin conductance response (SCR), electrocardiography (ECG), electromyogram (EMG) etc. as techniques to collect sensory data that can be correlated with psychophysica l load associated with driving. One of the concerns in correlating physiological measures with workload is that the sensory responses from the human body are due to many complex causal factors and driving workload alone cannot be considered as the only maj or factor.

PAGE 40

40 Measuring S ituational A wareness Situational awareness (SA) i s the degree to which a driver is aware of the surrounding environment, and specifically those elements that are relevant for the driving task (Endsley, 1995) There are three differe nt levels of SA: 1. SA1: Perception of the Situation (Corresponds to Control in Figure 2 4) 2. SA2: Comprehension of the Situation (Corresponds to Maneuver in Figure 2 4) 3. SA3: Projecting the situation into the future (Corresponds to Strategy in Figure 2 4) U sing driving simulators, situational awareness has been measured by a simulation freeze technique suggested by Endsley ( 1995 ) called Situational Awareness Global Ass essment Technique (SAGAT). In this technique, the simulator is stopped at pre decided intervals and subjects are asked questions corresponding to the three levels of SA SAGAT scores are calculated based on the percentage of correct answers. There are conc erns with the freeze probe techniques being obtrusive but SAGAT outperforms non freeze techniques in the extent of usage and validation. An appendix containing SA questions for all three levels are included at the end of this document. S ummary Of WL And S A Measurement S tudies A summary of the studies found on Driver Workload (WL) and Situational Awareness (SA ) is presented in Figure 2 10 details and the number of factors used in the experiment design. Most of the studies collect WL, while two studies collect both WL and SA in some form (2 2). NASA TLX and Peripheral Detection Task (PDT) are the most used measurement methods for WL, whereas SAGAT is used for measuring SA.

PAGE 41

41 Study WL SA Participants Factors (# of levels) Hancock et al (1990) NASA TLX, SWAT N=18, Age = 30 (21yr to 50yr) Maneuver (3) Verwey (2000) PDT, Audio Detection N = 24, 24 Age 30, 69 Age (2), TOD (2), Familiarity (2) Bolstad (2001) SAGAT N = 16x3 = 48 Age = 19,45,70 Age (3), Complexity (2) Hancock et al (2003) Break response time N = 19, 17 Age = 30, 60 Age (2), Distractor (2) Jahn et al (2005) PDT, ECG, NASA TLX N= 49, Age = 41.2 (SD=9.4) Display Size (2), Demand (2) Ma and Kaber (2005) PDT SAGAT N= 18, Age = 26.6 (SD=6.1) Cruise Control (2), Cellphone (2) Patten et al (2005) NASA RTLX, PDT N = 40, 39 Age= 39, 32 Mileage (2), Comp l exity (3) Kass et al (2006) Turns Missed SAGAT N = 24, 25 Age=15,19 Experience (2), Cellphone (2) Baumann (2007) Hit rate SAGAT N= 19, Age = 23.9 (SD=2.8) Locations (4), Tasks (4) Baldauf et al (2009) SWAT, EDA, Time Perception N=16, Age =34 (20yr to 54yr) City (1), Straight Road (1), Oncoming Traffic (1) Cantin et al (2009) Reaction Time N = 10x2 = 20 Age = 24,69 Age (2), Complexity (2) Filtness et al (2012) NASA TLX, PDT N = 20x2 = 40 Age = 27,37 Experience (2), Alcohol Dose (3) Arien et al (2013) PDT N=46, Age = 45.3 (20yr to 60yr) Curves (2), Gates (2) (Traffic Calming) Jansen et al (2016) ISA, NASA TLX N= 38, Age = 19.9 (SD=4.4) Task Demand (3) Figure 2 10 Summary of the Studies with Driver WL and SA Measurement Unless they use age or experience as a factor, most studies include only younger college

PAGE 42

42 ages are skewed towards extremes (too old or too young) as they address issues related to age or inexperience. The studies listed in Figure 2 10 consider WL and SA from perspectives such as improving traffic safety (traffic calming, alcohol usage, distraction from cellphones etc.), accommodating elderly (age), training the novice (experience) etc. From a traffic operations perspective, it is nece ssary to understand the relationship between driver performance and traffic conditions/facilities for a more typical set of drivers. Measuring D river C haracteristics Driver characteristics include a broad range of variables helpful in assessing the ability and nature of the driver. These can be measured by either observing or asking the drivers about their response to the situations. Each method with their pros and cons are explained in the next sub sections. Instrumented v ehicles Instrumented Vehic les are equipped with devices measuring speed, acceleration, and distance to the lead or lag vehicle, as well as eye tracking. For example, the 100 car naturalistic study ( Dingus et al, 2006 ) collected pre crash naturalistic driving data Recent work ( Kondyli et al, 2011 and Soria et al 2014 ) collected instrumented vehicle data to evaluate car following and lane changing behavior, and developed mo dels for enhancing simulation modeling. They also used the their aggressiveness. Focus g roup s tudy These have been helpful in understanding the driving tasks as well as the variability in factors associated with driving such as experience, awareness, risk taking attitude and motivation for compliance with traffic rules. Kondyli and Elefteriadou ( 2011 ) conducted focus groups to

PAGE 43

43 thinking process when driving along a freeway ramp merging segment. Sun and Elefteriadou ( 2012 ) conducted focus groups to identify driver types and study factors affecting lane changing. They concluded that four driver types could adequately represent t he observed heterogeneity in their sample. Driver b ehavior q uestionnaire (DBQ) Driver Behavior Questionnaire was first prepared by Reason et al. ( 1990 ) and it is known d on this study to address their needs. Li ( 2014 ) compared answers given in the Driver Behavior Questionnaire (DBQ) with actual behavior observed during an in car driving test. She categorized driving behavior into specific driving maneuvers and framed 2014 ) verified the reliability of answers given in DBQ to the respective driving behavior. Discussion This chapter provided a brief review of evolution of car following models in traffic engineering, followed by a section on popular microsimulators and behavior related parame ters used in them While the car following models have been constantly modified to produce reliable traffic operational measures, none of the classical or modern CF models incorporate driver behavior satis factorily. Driver Behavior models from the field of psychology are reviewed and common behavioral elements (workload (WL) situational awareness (SA) driver characteristics (DC) etc.) are identified along with their data collection techniques. It was observed that there are no set standards or widely accepted data collection methodology to collect the identified behavioral

PAGE 44

44 elements (WL, SA, DC ) This is mostly due to disconnect between traffic psychology and traditional traffic engineering. Based on th e information available in the literature, a framework is presented in the next chapter, followed by the respective data collection and data analysis.

PAGE 45

45 CHAPTER 3 METHODOLOGY AND DATA COLLECTION Figure 3 1 presents a framework relating driver characteristics and performance to traffic operations and highway design. This framework is used as the basis for this dissertation. The lower part of the figure presents microscopic elements of the system, while the top presents macroscopic elements. The solid lines represent the models and relationships that are established in the field of traffic operations. The dashed lines represent the models and relationships being studied in the field of psychology; their usage is virtually non existent in traffic operational analyses. Figure 3 1 Elements of Traffic Operations

PAGE 46

46 following, lane changing, gap acceptance etc. (Manjunatha et al, 2017). Traffic microsimulation tools use such models to predict prevailing traffic conditions at corridor or intersection levels (Manjunatha et al, 2013). Driver behavior and driver variability are represented by categorizing drivers into groups. H owever, there is no quantitative basis for establishing these groups At the same time WL and SA are not considered in the analysis. In the field of psychology, driver states (WL, SA etc.) are studied as a function of: 1 Design of the transportation facilities i.e. type of facility like signalized intersection, roundabout etc. (Verwey, 2000), type of geometric design (Arien et al, 2013) etc. 2 Prevailing traffic conditions (Patten et al, 2006) and complexities (Jahn et al, 2005) 3 Driver Characteristics such as age (Bolstad, 2001), experience (Patten et al, 2006), etc. 4 Level of automation (Ma and Kaber, 2005) etc. dynamically over time depen ding on the internal emotional factors and external environmental ph The way in which driver states and driver behavior are defined in psychology has not been carried over to how driver groups and categories are formed in traffic operational m odels. There have been some studies which propose a model (Task Capability Interface (TCI), Fuller, 2005), a construct (Drivability, Bekiaris et al, 2003), or a theoretical framework (TCI

PAGE 47

47 Implementation, Hoogendoorn et al, 2013) to address this issue, howe ver there is a lack of data driven studies to validate them. With innovations such as autonomous and connected vehicles, it is important to understand these missing links (Endsley, 2017) and create a data driven framework to address those (Parasuraman et al, 2008). The benefit of such a framework would be two fold; first, they would improve our prediction of various traffic phenomena such as breakdown (i.e., congestion triggering), which are highly dependent on human behavior; second, they would result i n improving human performance with better highway and transportation system designs. Thence the focus of this dissertation will be to find: 1. How prevailing traffic conditions affect driver states (i.e. WL and SA) 2. How driver states (WL and SA) affect d river behavior (CF and LC) 3. How driver characteristics (DC) affect driver behavior (CF and LC) The following section presents the methodology used for measurement of WL and SA, followed by the respective data analysis in order to develop a (Chapter 5), (Chapter 6) and driver characteristics The use of a driving simulator provides an opportunity to study car following behavior in a controlled environment. While they present several advantages such as controllability, reproducibility, ease of data collection etc., the driving simulator experiments need to be designed carefully to achieve a reliable behavioral fidelity ( d e Winter et al, 2012). Hence, a driving simulator experiment is designed to collect WL, SA and DBQ data along with vehicle trajectories T wo control scenarios consisting of no ambient traffic were designed to establish base conditions as well as to provide data for develo ping a novel estimation method where the

PAGE 48

48 parameters are estimated in order to replicate the specific behavior, as opposed to reducing errors of aggregate measures. Following this, Driver Workload (WL) and Situational Aw areness (SA) are measured for the drivers at three levels of traffic densities Low, Medium, and High and at three different spatial segments of the freeway Basic, Approach and Merge segments Peripheral Detection Task (PDT) and NASA TLX questionnaires were employed to measure Workload (WL), whereas SAGAT questionnaire was developed to measure Situational Awareness (SA). Driver Behavior Questionnaire (DBQ) was employed at the end of the experiment to collect information on Driver Characteristics (DC). The details of the experiment design, drivin g simulator used, participants involved and the driving scenarios employed are discussed in this chapter. Experiment E quipment The driving simulator (Figure 3 1 ) is a full car cab (4 door sedan) with seven visual channels. The three forward channels create a 180 degree field of view (FOV). This wide FOV is accomplished by connecting three flat screens with scenes provided by three high resolution projectors. The rear scene is also projected on a flat screen and viewed through the in cab rear view mi rror. The side view mirrors and a virtual dash are simulated with LCD panels. Figure 3 2 Driving Simulator

PAGE 49

49 Participants A pre screening questionnaire (Appendix B) made in compliance with Institutional Review Board (IRB 201700047 ) was used to recruit f orty participants (20 male and 20 female) between the ages 18 and 60 All participants had an active driving license with a minimum driving experience of one year. The average age of the participants is 40.6 years and the standard deviation was 11.2 years. The demographic split can be seen in Figure 3 3 Figure 3 3 Gender Age Mix in the group of participants The prescreening questionnaire and the distribution of drivers by their occupation, experience, aggressiveness etc. can be found in the Appendix. Experimental D esign And D riving S cenarios A practice scenario was designed for participants to get acclimated to the simulator environment. Participants would generally take between 4 to 6 minutes for a satisfactory initial acclimation. The Kennedy Simulator Sickness Q uestionnaire (SSQ) ( Kennedy et al, 1993 ) was administered to check for simulator sickness (Appendix C) Participants were then asked to drive in two control scenarios: 1. A scenario on a straight, empty highway with varying speed limits. Their speed limit compliance was recorded to use it as a surrogate for driver aggressiveness. 0 2 4 6 8 10 12 14 20s 30s 40s 50s Female Male

PAGE 50

50 2. A scenari o in which the participants were required to follow a lead car which was scripted to travel straight at a constant speed under the speed limit. Except for the subject and the lead vehicles, there was no other traffic on the road. Hence the following patter ns observed were free from the influence of other ambient traffic. The trajectories of the following process were recorded at three different lead vehicle speeds. Next, Driver Workload (WL) and Situational Awareness (SA) were measured for the drivers at th ree levels of traffic densities Low, Medium, and High ( 3 1) and at three different spatial segments of the freeway Basic, Approach and Merge segments (Figure 3 3) Table 3 1 Traffic States Traffic State Density (veh/mi/ln) Traffic flow (veh/hr/ln) Average Speed (mph) LOS (HCM) Low 10.7 (0.5) 798 (26) 74 (1.4) A/B Medium 18.6 (0.6) 1309 (66) 70 (1.7) B/C High 31.0 (0.7) 1961 (52) 63 (2.8) D S tandard deviations are provided in the parenthesis Figure 3 4 Measurement Sections For Workload, Peripheral Detection Task (PDT) as explained in Chapter 2 was employed (Figure 3 4). Using the feedback buttons inside the car, the participants were asked to respond to

PAGE 51

51 randomized onscreen visual cues while driving on a freeway. At the end of each PDT d rive, NASA TLX (as explained in Chapter 2) was administered The rating scale and the explanation of the method is included in a ppendix D Following this, the participants were asked to drive on the freeway section ti ll their screen faded away at a random section (basic, approach or merge) Once the screen faded out, the simulation would be stopped temporarily and Situational Awareness questions were asked. A small break (1 to 2 minutes) is given before starting each s cenario, participant feedback is considered and SSQ is employed whenever necessary. Figure 3 5 Driving Scenarios on a Freeway for a given density The order in which the traffic densities (Low, Medium, High) were introduced and the segments at which the screen faded out for SAGAT questions (Situational Awareness Global Assessment Test) were randomized. This was done for several reasons: To avoid par ticipants finding set patterns and adjust accordingly To avoid the effects of familiarity and fatigue biasing the results An example of SAGAT questions corresponding to three different levels of SA: SA 1 (Perception): Are you traveling above or below the speed limit?

PAGE 52

52 SA 2 (Comprehension): What is the relative speed of the traffic compared to y SA 3 (Projection): Are you losing or gaining on the vehicle in front (in the same lane)? Complete set of SAGAT questions can be found in the appen dix E After the participants completed the driving they were asked to fill out responses to a Driver Behavior Questionnaire (DBQ appendix F ) In summary t he following driver behavior data are obtained from the driving simulator experiment: 1. Driver characteristics from the pre screening form 2. Trajectories of the control scenarios (Used for developing a novel estimation methodology) 3. Workload from PDT (for 3 densities x 3 segments = 9 com binations) 4. Workload from NASA TLX (Along with individual NASA TLX categories) 5. Situational Awareness from SAGAT (for 3 densities x 3 segments = 9 combinations) 6. Driver Behavior Questionnaire (DBQ) responses 7. Trajectories of the scenarios driven In the next chapter, a methodology is presented to take advantage of the car following data from the control conditions to estimate Wiedemann car following parameters. This is followed by preliminary results.

PAGE 53

53 CHAPTER 4 A NOVEL METHOD FOR ESTIMATION OF CAR FOLLOWING PARAMETERS The conventional methods of calibrati ng car following parameters are based on minimizing the mean square error between observed and simulated aggregate traffic measures such as average speed, travel time and delay etc (Ma njunatha et al, 2013) However, different vehicle trajectories can result in the same aggregate measure. Hence these methods cannot ensure that the properties of the observed and simulated vehicle trajectories are the same (Menneni et al, 2008). Though so me studies use naturalistic data (Durrani et al, 2016, Higgs et al, 2011) to minimize the error between observed and simulated trajectories to find car following parameters, they employ the same heuristic methods such as genetic algorithm used by conventional methods. Parameter values calibrated using such methods may match the sample trajectories, however they do not reflect the physical meaning/behavior of the specific parameters. Detailed discussion on trajectory based calibration settings (i.e., combinations of algorithms, measures of performance, and goodness of fit functio ns) can be found in Punzo et al ( 2012 ). The use of a driving simulator provides an opportunity to study car following behavior in a controlled environment. The car following patterns can be clearly recorded through driving simulators by fixing the lead veh icle speed to a constant value and allowing the subject drivers to frequently enter the following regime. The following sections describe the relationship between the Wiedemann parameters and thresholds and explain how the parameters can be esti mated to match the regime thresholds themselves as opposed to trajectories or aggregate measures.

PAGE 54

54 Relationship Between Wiedemann Parameters And Regime Thresholds (Aghabayk et al, 2013) The values can be seen in Figure 4 1 along with definitions of the Wiedemann parameters (CC0 to CC9) as given in VISSIM documentation. Among the se parameters, CCO to CC6 affect car following directly. Figure 4 1 Wiedemann parameters and regimes plot used in VISSIM T he following equations (4 1 to 4 6) from Aghabayk et al. ( 2013) can be used to study the relationship between the Wiedemann 99 parameters and thresholds. These equations have been developed for a specific case of an individual driver, who is following a lead vehicle traveling at a fixed speed. Since these equations ar e for a specific event, there are no random parameters. AX is the minimum stationary gap: ( 4 1) AX = L + CC0 Where L is length of the lead vehicle. BX is the speed dependent safe gap: ( 4 2) BX = AX + (CC1 x v) Where v i s the velocity of the subject vehicle ( 4 3)

PAGE 55

55 SDX = BX + CC2 SDV is the perception point while approaching from long distances: (4 4 ) CLDV is the perception point while approaching from short distances: ( 4 5) OPDV is the perception point while lagging at short distances: ( 4 6) Measurement M ethodology The regime plot shown in Figure 4 1 represents the original Wiedemann model ( Wiedemann, 1974 Aghabayk et al, 2013 ) representing equations 4 1 to 4 6 is shown in Figure 4 2. The thresholds in Figure 4 2 have been drawn to scale using VISSIM defaults for a constant lead vehicle speed of 40 mph (~18m/s) and compared with plots from an actual trajectory data for the same speed range. The x axis represent s the speed difference whereas the y axis is the distance between the subject and lead vehicles. Figure 4 2 Trajectory and thresholds as per default VISSIM values v/s actual trajectory space headway in meters (m) within which the subject vehicle is in the car following regime.

PAGE 56

56 in m/s within the following zone. CC3 is the time interval (s) between the time of perception of a slower lead vehicle to the 1 represents the rate of approach at lower distances. Geometrically, the thresholds define a fou r 4 3). The top and bottom are bounded by the straight lines SDX and BX, with CC2 being the height of the ceiling (Eqn. 4 2). Figure 4 3 The left and the right si des are bound by OPDV and SDV respectively. While OPDV and CLDV are second degree curves with CC6 defining their curvature (Eqns. 4 5 and 4 6), SDV is a straight line with CC4 as the intercept at SDX and CC3 as its slope (Eqn. 4 4). In the control driving scenario of the driving simulator experiment there are no other vehicles on the road other than the lead and subject vehicles. The lead vehicle travels at a constant speed and the subject drivers are asked to follow the lead vehicle so that clear followin g patterns can be recorded (Figure 4 2 ). Using the physical and geometrical relationships established (Figure 4 3) scripts are written to extract thresholds for individual drivers from the trajectory data. Using equations 4 1 to 4 6 and the extracted thre sholds, the Wiedemann

PAGE 57

57 parameters (CC1 to CC5) are calculated at different speeds for a given driver. In the next section, the measurement results are presented. Estimation R esults To illustrate the data analysis process, the calculated values of the Wiedemann parameters for a given driver (Driver 12) are provided in 1. VISSIM default values, which were calibrated upon studies in Germany has also been provided for comparison Table 4 1 Wiedemann parameters at different speeds for Driver 12 Driver 12 CC1 (s) CC2 (m) CC3 (s) CC4 (m/s) CC5 (m/s) 20 mph 0.63 6.99 4.84 0.93 0.78 40 mph 0.57 0.56 7.16 0.34 0.28 60 mph 0.31 6.65 11.05 0.81 0.87 S tandard D eviation 0.14 2.95 2.56 0.25 0.26 VISSIM Defaults 0.9 4 8 0.35 0.35 Mean 0.50 4.74 7.68 0.69 0.64 Like 4 1, these parameters are estimated for all the forty drivers at three different lead vehicle speeds and the summary is provided in Figure 4 4. The last row indicates whether there is a statistically significant difference (at 95% confidence) between the distributions of the fitted values compared to the VISSIM defaults. Except for CC3, all fitted values differ significantl y from VISSIM defaults. The value of safe minimum time headway (CC1) increases with speed, suggesting that the safe time headway is not a constant and hence safe space headways may be underestimated when calculated as a linear function of speed (Eqn. 4 2). The range of oscillation in terms of distance (CC3) and speed (CC4, CC5) is smallest for 40 mph, i.e. estimating/matching the speed of the lead vehicle is better at mid range speeds. The increases with the speed.

PAGE 58

58 Figure 4 4 Wiedemann parameters estimated for different speeds for all drivers Now that it is established that the overall estima ted mean values vary significantly from VISSIM defaults, in the following sections the comparisons are made between sub groups mean values and the overall mean values for a given parameter. Estimation of the Wiedemann parameters by age group is shown in Figure 4 5 The last column in each group indicates whether there is a statistically significant difference (at 95% confidence) between the mean of the age group compared to that of the overall mean of the given parameter. Though there is no clear trend in paramet er values by age group ( Figure 4 5 ), 3 out of 5 parameters in each age group differ significantly from their overall means. Figure 4 5 Wiedemann parameters by age group In one of the control scenarios, participants were asked to drive on an empty highway with varying speed limits. Their speed limit compliance is recorded to be used as a surrogate for driver aggressiveness. Participants who travel within +/ 5% of the speed limit are considered as respectively. Age 20s Sig 30s Sig 40s Sig 50s Sig All Drivers Mean SD Dif? Mean SD Dif? Mean SD Dif? Mean SD Dif? Mean SD CC1 1.26 0.32 Y 2.06 0.41 Y 1.60 0.37 N 1.33 0.52 N 1.60 0.21 CC2 3.26 0.75 Y 4.01 1.20 N 4.35 0.56 Y 3.03 0.92 Y 3.74 0.46 CC3 8.60 0.98 Y 7.58 0.74 Y 8.14 1.23 N 7.57 0.81 N 8.10 0.60 CC4 0.79 0.25 N 0.60 0.11 Y 1.02 0.23 Y 0.57 0.14 Y 0.78 0.12 CC5 0.64 0.13 N 0.64 0.12 N 0.58 0.10 Y 0.79 0.22 Y 0.65 0.08

PAGE 59

59 Estimation of the Wiedemann parameters by their aggressiveness is provided in Figure 4 6 The last column in each group indicates whether there is a statistically significant difference (at 95% confidence) between the mean of the group compared to that of the overall mean of the given parameter. Behavior Conservative Sig Moderate Sig Aggressive Sig All Drivers Mean SD Dif? Mean SD Dif? Mean SD Dif? Mean SD CC1 (s) 2.04 0.51 YES 1.57 0.41 NO 1.27 0.37 YES 1.60 0.21 CC2 (m) 3.01 0.90 YES 3.96 1.20 NO 3.98 0.56 NO 3.74 0.46 CC3 (s) 9.67 1.49 YES 7.85 0.74 NO 7.27 1.23 YES 8.10 0.60 CC4 (m/s) 0.67 0.20 NO 1.04 0.11 YES 0.53 0.23 YES 0.78 0.12 CC5 (m/s) 0.51 0.10 YES 0.70 0.12 NO 0.69 0.10 NO 0.65 0.08 Figure 4 6 Wi e demann p arameters b y driver aggressiveness There is a clear trend in safe headway (CC1) values: aggressive drivers tend to maintain lower headways than conservative drivers. Similarly, they also take less time from perception of meter values for moderate drivers match closely with those of the overall means. Whereas for aggressive and conservative drivers there are many significant differences. Estimation D iscussion A driving simulator experiment is devised to collect trajectori es of car following z estimate the Wi e demann parameters for the collected data at different speed ranges for individual drivers. The results are summarized and presented in table s by speed range, driver age, and driver behavior respectively. The results show that the Wi e demann thresholds vary over both speed ranges and driver groups. Except CC3, all other estimated Wiedemann parameters showed statistically significant deviation from the VISSIM defaults. W hen compared to the overall estimated parameter set only the parameters obtained for the moderate group of drivers matched closely. Also, the mean and s tandard deviations of each

PAGE 60

60 behavior sub group were significantly different from those of the overall distribution. This indicates using one distribution to represent a wide range of behaviors may not be sufficient. The behavior recorded in this study fo cuses on the car following under the influence of the lead vehicle only (i.e surrounding traffic is absent). This study case can be us ed as base condition to compare the effect of surrounding traffic, geometrical elements etc. on the car following behavio r. Carefully designing the driving simulator experiment and calibrating the parameters to match the thresholds obtained, results in a better model/ estimation method. However, a single set of parameters still does not adequately cover the range of behaviors exhibited by the drivers.

PAGE 61

61 CHAPTER 5 DRIVER STATES: WORKLOAD AND SITUATIONAL AWARENESS dynamically over time depending on the internal emotional factors and external environmental factors. The impact of change in driver states on traffic operations occurs The overall objective of this task is to measure WL and SA for different drivers and as a function of freeway conditions, and develop a framework for considering the effects of driver performance and their incorporation into traffic analysis models. To achieve thi s, th is chapter aims to: 1. E xamine the relationship between WL and SA for freeway traffic c onditions 2. Compare two different measures of Workload (NASA TLX and PDT) 3. Categorize drivers into groups based on WL and SA, and examine driver characteristics and behavior across freeway conditions Based on the objectives above, the following Research Questions (RQ) will be addressed: RQ1. Do different types of freeway sections influence driver states (WL and SA) differently, and if yes, how? RQ2. Do changes in traffic density influence driver states (WL and SA) and if yes, how? RQ3. Do drivers use compensatory mechanisms for ??? and what is the effect of such compensation? RQ4. Are different methods of WL measurement (NASA TLX and PDT) consistent? RQ5. Can drivers be classified based on WL and SA states, and if so what are the properties of these classes?

PAGE 62

62 RQ6. How does the relationship between WL & SA vary across freeway conditions for different drivers? RQ1 Do Types Of Freeway Sections Influence Driver States ( WL And SA ) And If Yes, How? Figure 5 1 show s the variation of SA and WL (PDT) by freeway segment. Following Situational Awareness There was a significant increase (p<0.05) in SA (Figure 5 1a) from Basic (Mean 60%, SD 14%) to Approach segment (Mean 72%, SD 17%). However, SA remained marginally the same from Approach (Mean 72%, SD 17%) to Merge segment (Mean 70%, SD 16%). Workload (PDT) Despite the increase in SA, Workload (PDT) levels remained the same (Figure 5 1b) from Basic (Mean 35%, SD 16%) to Approach Segment (Mean 34%, SD 18%). However, Workload (PDT) increased significantly (p<0.1) from Approach (Mean 34%, SD 18%) to Merge Segment (Mean 41%, SD 15%). Hence it can be concluded that the WL and SA are influenced by the typ e of freeway segment. (a) (b) Figure 5 1 Driver States by Type of Freeway Segment 60% 72% 70% 0% 20% 40% 60% 80% 100% Basic Approach Merge SA by Freeway Segment 35% 34% 41% 0% 20% 40% 60% 80% 100% Basic Approach Merge WL (PDT) by Freeway Segment

PAGE 63

63 RQ2 Do Changes In Traffic Density Influence Driver States ( WL And SA ) And If Yes, How? Figure 5 2 shows the variation of SA and WL (PDT) by traffic density. Following Situational Awareness There was a significant increase (p<0.05) in SA (Figure 5 2 a) from Low (Mean 61%, SD 16%) to Medium density (Mean 71%, SD 18%). However, SA remained marginally the same from Medium (Mean 71%, SD 18%) to High density (Mean 70%, SD 17%). Workload (PDT) Despite the increase in SA, Workload (PDT) levels remained the same (Figure 5 2b) from Low (Me an 32%, SD 16%) to Medium density (Mean 34%, SD 20%). However, Workload (PDT) increased significantly (p<0.05) from Medium (Mean 34%, SD 18%) to High Density (Mean 45%, SD 21%). Hence it can be concluded that WL and SA are influenced by the level of traffi c density. (a) (b) F igure 5 2 Driver States by Level of Traffic Density 61% 71% 70% 0% 20% 40% 60% 80% 100% Low Medium High SA by Traffic Density 32% 34% 45% 0% 20% 40% 60% 80% 100% Low Medium High WL (PDT) by Traffic Density

PAGE 64

64 RQ 3 Do Drivers Use Compensatory Mechanisms And What Is The Effect Of Such Compensation? Figure 5 3 shows the variation of mean speed of drivers by type of freeway segment and traffic density. The graph for mean speed is a mirror image of workload, indicating that as the workload increases, the drivers try to decrease it by decelerating to lower speeds as a compensatory mechanism. The differences in mean speed between each freeway segment as well as the level of traffic density were found to be statistically significant (p<0.05) for all the combinations. Though WL remains the same between basic and appr oach segments as well as between low and medium densities, the mean speed decreases. However, for merge segments and high traffic density, despite the decrease in mean speeds, WL increases (Figures 5 1 b and 5 2 b). With these observations, one could argue t hat the drivers tend to lower their speed to maintain a segment and high traffic density) WL inevitably increases. (a) (b) Figure 5 3 Mean Speed as an Indicator of Compensatory Mechanism One could argue that driver speeds at high density could be physically limited by surrounding vehicles. However, the mean speed at high density (51 mph) in Figure 5 3 b is significantly (p<0.05) lower than the speed of surrounding simulated drivers (63 mph) who are 68 65 60 0 20 40 60 80 100 Basic Approach Merge Mean Speed (mph) by Freeway Segment 73 68 51 0 20 40 60 80 100 Low Medium High Mean Speed (mph) by Traffic Density

PAGE 65

65 modeled through an analytical formula ( 3 1 ). This observation confirms that the decrease in speeds is indeed because the subjects chose to decelerate, as opp osed to physical space c onstraints. It also reiterates the need to model human behavior appropriately to capture real life phenomena such as traffic breakdowns (Manjunatha et al, 2017). Hence it can be concluded that drivers use compensatory mechanisms su ch as decelerating to a more comforspeed, to achieve an acceptable/desirable level of WL. R Q 4 Are Different Methods Of Workload Measurement (N ASA T LX And P DT ) Consistent? NASA TLX method (Hart et al, 1988) is a questionnaire method where drivers are asked to rate the driving difficulty on a 20 point scale in six categories: Mental Demand: How much mental and perceptual activity was required? Were the driving and PDT task easy or demanding, simple or complex? Physical Demand: How much physical act ivity was required? Were the driving and PDT task easy or demanding, slack or strenuous? Temporal Demand: How much time pressure did you feel due to the pace at which the PDT tasks or task elements occurred? Was the pace slow or rapid? Performance: How successful were you in performing in terms of driving and PDT tasks? How satisfied were you with your performance? Frustration Level: How irritated, stressed, and annoyed versus content, relaxed, and complacent did you feel during the driving and PDT tasks ? Effort: How hard did you have to work (mentally and physically) to accomplish your level of performance?

PAGE 66

66 This is followed by asking drivers of their perceived importance of each category through pairwise comparison to determine weights of each category a s given in 5 1 Higher weights suggest higher relative importance according to the participants. Table 5 1 Average weights calculated from NASA TLX self rating responses Category Mental Physical Temporal Performance Effort Frustration Weight 24% 11% 17% 21% 17% 9% As shown, Mental Demand and Performance are weighed heavily i.e. rated more important by the drivers, followed by Temporal Demand and Effort respectively. Physical Demand and Frustration are weighted the least i.e. the participants think that physical requirement of the driving was r elatively low and the driving scenarios caused relatively less frustration, hence their impact on overall Workload was less compared to other categories. The average ratings in each category can be seen in Figure 5 4 In all categories, high density has th e highest ratings. Hence higher WL could be attributed to higher traffic density. Figure 5 4 NASA TLX Self Rating Category Averages 58 39 54 31 49 28 61 39 52 27 54 28 70 47 57 33 59 41 0 10 20 30 40 50 60 70 80 90 100 Mental Physical Temporal Performance Effort Frustration NASA TLX Workload Components (%) Low Medium High

PAGE 67

67 The category ratings increase or stay almost the same between low and medium densities. medium density condition is better than that of low or high densities. This agrees well with the theory (de Waard, 1996) that there is a range of task demand (in this case traffic density) where the performance is optimal (Note: Performance is on a reverse scale i.e. smaller the better). However, for high traffic density, all the rating categories increased. Largest increases were for Frustration and Mental demand respectively. (a) (b) Figure 5 5 Relationship between WL (PDT) and WL (NASA TLX) measures 32% 34% 45% 46% 47% 55% 0% 20% 40% 60% 80% 100% Low Medium High WL by Traffic Density WL (PDT) WL (NASA-TLX)

PAGE 68

68 When the averages of WL (PDT) and WL (NASA TLX) are compared, they show a similar trend to chang es in traffic density (Figure 5 5 a). The low and medium densities have similar WL whereas there is a significant increase in WL for high traffic density. Howev er, when WL (PDT) and WL (NASA TLX) are compared at the individual level, they do not show a strong correlation. Figure 5 5 b is a plot of the PDT vs. the NASA TLX and NASA was found to be 0.02 which suggested a weak correlation. As mentioned in Jahn et al. (2005), the PDT measure provides a reasonable bandwidth to measure inter indiv idual differences, whereas NASA TLX is helpful in understanding the effect of experiment manipulation on the participants. The increase in Mental and Physical Demands (in NASA TLX) from Low to High traffic densities confirm the same for this study (Figure 5 4). Hence it can be concluded that NASA TLX and PDT measures of workload are not consistent at the individual level, however they show similar trends when aggregated. NASA TLX sub categories are also useful in understanding the effects of design and traf fic changes on drivers. While PDT measure and SA are useful in understanding what the driver experience, NASA TLX categories (Mental Demand, Physical Demand etc.) are useful in understanding how driver think. RQ 5 Can Drivers Be Classified Based On WL An d SA States, If So What Are The Properties Of These Classes? A simple two step clustering was conducted and the drivers were classified into two groups based on their mean WL (PDT) and SA measurements across different conditions. The number of clusters wer e chosen based on the Silhouette measure of cohesion and separation as well as the ratio of largest cluster to the smallest ( Figure 5 6 ).

PAGE 69

69 Number of Clusters (k) Average Silhou e tte Width Size of the Smallest Cluster (S) Size of the Largest Cluster (L) L/S 1 0 0 2 0.6 18 22 1.2 3 0.6 5 22 4.4 4 0.5 5 18 3.6 Figure 5 6 Clustering Statistics Figure 5 7 shows mean WL (PDT) and SA measurements. Each data point corresponds to mean values for individual drivers. The two groups of drivers are identified by hollow (Group A) and solid (Group B) circles. The reference lines are drawn at mean va lues of WL and SA. On visual observation, most drivers in 2 st quadrant (top left) belong to Group A, whereas all drivers in 4 th quadrant (bottom right) belong to Group B. Group A drivers generally have lower SA and experience higher WL compared to Group B. Figure 5 7 Mean WL (PDT) and SA for the clustered groups Figure 5 8 shows the properties of the driver classes based on the information collected in the pre screening questionnaire. Group A consists of somewhat older drivers and a higher percentage of females. They are less likely to drive on a regular basis than group B and more

PAGE 70

70 likely to use GPS (Figure 5 8 ). The driving simulator did not have GPS, hence lower SA from group A could be explained with this. Group B drivers rate themselves more aggressive and more likely to enjoy driving. Property Group A Group B Comments N 18 22 WL (PDT) 49.2% 26.3% Measured SA 60.5% 73.1% Measured Age 42.7 36.8 Gender 39% 61% 59% 41% % Male % Female Frequency of Driving 2.2 2.5 1= <4 hr/week 4 = >14hr/week GPS usage 2.2 2 1= Always, 4 = Never Enjoy Driving 7.1 7.4 10 point scale Aggressive 4.1 5.4 10 point scale Tickets 5.50% 22.70% At least one in the past 1 year Involved in Accidents 33.30% 27.30% At least one in the past 5 years Figure 5 8 Properties of driver classes based on pre screening questionnaire Drivers from Group B are also more likely to get traffic violation tickets. However, they are less likely to be involved in accidents than Group A (Figure 5 8 ). To have a broader understanding of why Group B drivers have favorable driver states (High SA, Low WL) compared to Group A (Low SA, High WL), measures such as mean speeds, SA sub levels and NASA TLX categories are investiga ted by driver group in Figure 5 9 Figure 5 9 a provides SA sub levels by driver group. The largest difference is seen at Level 3 SA (Projection) i.e. Group B drivers are better at comprehending and projecting the immediate traffic consequences. An effect of this can be seen in th e NASA TLX components (Figure 5 9 c). Group B drivers assess average Mental Demand (66%) to be much higher than Group A (59%). Hence Group B increase their performance level (23%) much better than tho se of Group A (39%) (Note: Performance is on a reverse scale i.e. smaller the better). Group B drivers have to compensate for Lower SA, Performance and manage the higher WL, hence they use compensatory mechanisms and their mean speeds are lower than those of Group A acros s

PAGE 71

71 different densities (Figure 5 9 b). The difference in mean speeds the two driver groups is minimal for medium density (Figure 5 9 b), creating homogenous traffic conditions which are ideal for traffic operations. (a) (b) (c) Figure 5 9 Driver State Sub Levels by Driver Groups Hence it can be concluded that the drivers can be classified based on WL and SA states. The properties of each driver group is as listed in 5 3 NASA TLX categories a re useful in 75 68 39 86 76 57 0 20 40 60 80 100 SA1 SA2 SA3 SA Levels A B 70 67 49 76 69 53 30 40 50 60 70 80 Low Medium High Mean Speed (mph) Mean Speed (mph) v/s Traffic Density A B 59 40 56 39 55 35 66 43 52 23 54 30 0 10 20 30 40 50 60 70 80 90 100 Mental Physical Temporal Performance Effort Frustration NASA TLX Workload Components (%) Group A Group B

PAGE 72

72 RQ 6 How Does The Relationship Between WL & SA Vary Across Freeway Conditions For Different Driver Classes ? Figure 5 10 shows the relationship between WL and SA across different conditions. Each data point corresponds to mean values for an individual driver. The two groups of drivers are identified by hollow (Group A) and solid (Group B) circles. The reference lines are drawn at mean values of WL and SA for the specific area represent centroids of each sub group. The top left (Low SA, High WL) quadrant represents the worst driver state in terms of safety and operation. Drivers on the top right (High SA, High WL) are situationally aware and are likely to drive safely, however from an operations perspective, they may not make the best decisions in terms of car following or lane changing. The bottom left (Low SA, Low WL) represents low WL conditions, however safety could be an issue for drivers in this zone due to low SA. The bottom right (High SA, Low WL) represents t he best conditions from both safety and operations perspective. The goal of designing and operating highway facilities should be to ideally get the majority of drivers into this zone (High SA, Low WL). i n terms of WL and SA (Figure 5 10 Similarly, within different levels of in terms of WL and SA (Figure 5 10b ). Hence from a traffic operations perspective, maintaining the traffic at medium density would be best for overall system performance. This is consistent with the observ ation in literature that the accident rates are the lowest for medium density traffic overlap in driver states between the two driver groups.

PAGE 73

73 (a) WL and SA by Freeway Segment (b) WL and SA by Traffic Density Figure 5 10 Relationship between WL and SA across different conditions

PAGE 74

74 This indicates that the two driver groups are or the mental demand (Figure 5 8 ) are low. This is consistent with theories in traffic operations (Kondyli et al, 2013) and could explain how traffic breakdowns occur or why the freeway capacity is stoc hastic. At low and medium density traffic, driver behavior is homogenous, however as the traffic gets congested and approaches from medium to high density, the driver groups start responding differently and this heterogeneity leads to breakdown at varying flow and density levels. Hence it can be concluded that the nature of the relationship between WL and SA for the two driver groups follow similar trends across different traffic conditions. Group A drivers have lower SA and experience higher WL than group B drivers in all the conditions. The differences between the two groups is minimal for low density (resulting in homogenous traffic) and much larger for higher density (resultin g in heterogeneous traffic). Discussion On Driver States The following are the conclusions from this section: 1 Multivariate analysis revealed that the type of freeway segment as well as the level of traffic density both affect WL and SA. However, their interaction effect was found to be statistically insignificant. 2 Drivers decelerate to a comforspeed to cope with increased WL. The differences in mean speed between each freeway segment as well as level of traffic density were found to be statistically significant. This is likely one of the contributing factors in freeway merge breakdown (Kondyli et al, 2013). 3 Cluster analysis was used to classify drivers into groups based on SA and WL. It was concluded that the drivers can be classified into two groups: The first group of drivers had significantly lower SA and experienced higher WL than the second group. The WL and SA

PAGE 75

75 patterns for the two driver groups were found to be consistent across types of freeway segments and levels of traffic density. 4 The two driver groups experience similar WL for low density traffic, whereas they have similar SA and m ean Speed at medium density traffic (i.e. homogenous traffic conditions). The difference between the two groups was larger at high density (i.e. heterogeneous traffic conditions). Medium density has the best combination of WL (low) and SA (high) for traffi c operations. This is further supported in literature through low accident rates (Martin, 2002). Studying such interactions between the driver groups in conjunction with vehicles trajectory and other measurements has the potential to explain traffic phenom enon such as capacity drop and breakdowns. 5 Though WL (PDT) and WL (NASA TLX) showed similar aggregate trends, their measurements at the individual driver level did not show any correlation. However, the individual categories of the rating technique, su along with SA sub levels were useful in understanding the behavioral differences between the driver groups. The hypothesis of Figure 3 1 is that changing conditions (traffic density, type of segments) change driver cognitive states (WL, SA) which in turn influence driver behavior (CF, LC ) Answer s to the r esearch questions from this chapter have shown how the driver states i.e. WL and SA change across different freeway sections and under different levels of traffic density. Hence t he next step is to determine whether the change in conditions (traffic density, type of segments) change s driver behavior, if yes then how to model the relationship between driver states and driver behavior in a microsimulation context The f ollowing chapter explores the answer s to thes e question s

PAGE 76

76 CHAPTER 6 DRIVER BEHAVIOR: CAR FOLLOWING AND LANE CHANGING Behavior g : CF, LC cognitive response to change in driver states The objective of this section is to examine CF and LC related measures from the vehicle trajectories for different drivers and as a function of freeway conditions, and develop a framework for incorporating the effects of driver performance into traffic analysis models. To achieve this, t h is chapter aims to: 1 E xamine the CF and LC measures for freeway traffic c onditions 2 Compare the CF and LC measures between the driver groups based on WL an SA 3 Explore modeling the relationship between Driver States and Driver Behavior Based on the objectives above, the following Research Questions (RQ) will be addressed: RQ1. Do drivers classified based on WL and SA states, have different CF and LC behavior? RQ2. Do different types of freeway sections influence dr iver behavior (CF and LC) differently, and if yes, how? RQ3. Do changes in traffic density influence driver behavior (CF and LC) and if yes, how? RQ4: How do Driving Regimes and the Minimum Safe Headway Vary by Driver Group and Traffic Condition? RQ5: How to Model the Relationship between Driver States and Driver Behavior?

PAGE 77

77 RQ1 Do Drivers Classified Based On WL And SA States, Have Different CF Behavior ? Figure 6 1 shows the car following parameters estimated for the control scenario (i.e. following a car with no ambient traffic) by driver group. The last row indicates whether there was significant difference between the driver groups for the given CF parameter. Since there are no other external factors such as the type of segment or level of traffic density influencing the car following behavior, the differences here can be assumed to be only due to driver heterogeneity. Figure 6 1 Car Following Parameters (Estimated for Control Scenario) by Driver Group CC1 Minimum Safe Headway (s) : Group B drivers experiencing lesser WL and having higher SA manage to maintain significantly (p<0.05) smaller safe headways than Group A drivers. CC2 Following Variation (m) : Group B have smaller following variation (i.e. longitudinal oscillation in terms of distance) than Group A drivers. However, the difference is not statistically significant. CC3 Approach rate (s) : There is no significant difference between the two driver groups. CC 4 and CC5 Oscillation Limits (m /s ) : CC4 and CC5 have to be analyzed together since they represent the variation of speed difference ( v). While CC2 represents longitudinal oscillation in terms of distance, CC4 and CC5 represent longitudinal oscillation in terms of speed. Group B drivers oscillate significantly les s (p<0.05) than G roup A drivers. This demonstrates better longitudinal control during following from Group B (Low WL, High SA) drivers.

PAGE 78

78 Hence the car following parameters sho w that Group B drivers follow the leader vehicle closely (smaller CC1) with better longitudinal control (CC4, CC5). This is consistent with the theory that driver performance is better when WL is relatively l ower and SA is high er Traffic with higher proporti on of Group A drivers is likely to have lower capacities (due to larger CC1) and the oscillations are less likely to propagate for longer distances (due to smaller CC4). RQ2 Do Types Of Freeway Sections Influence Driver Behavior And If Yes, How? Figure 6 2 and Figure 6 3 show longitudina l and lateral behavior respectively, of the two driver groups under different freeway segments Longitudinal Behavior Headway (mean) is the average of the headway measured, irrespective of whether the subject was in the following state or not. There are two differences to note from Figure 6 2a : 1 Headways decrease significantly ( p <0.05) for both class es of drivers for the m erge segment 2 Headway s for group B drivers are significantly ( p <0.05) l ower than th ose of group A drivers for all type s of segments. This is consistent with the observation from control conditions (Figure 6 1) for CC1 (minimum safe headway during following) Similarly, there are two differences to note from Figure 6 2b : 1 Standard Deviation of Longitudinal Acceleration (SDLA) increases significantly ( p <0.05) for both class of drivers for m erge segment 2 Standard Deviation of Longitudinal Acceleration (SDLA) for group B drivers is significantly ( p <0.05) less than that of group A drivers for all type of segments. This is consistent with the observation from control conditions for CC4 and CC5 (Longitudinal Oscillation expressed in terms of speed difference) (Figure 6 1).

PAGE 79

79 (a) (b) Figure 6 2 Longitudinal Behavior by Freeway Segment Lateral Behavior Figure 6 3a provides the average number of lane changes by segment. Most of the lane changes for both driver groups occur at the approach segment. One possible reason for this is that the drivers who prefer to stay on the left l ane to avoid merging traffic, might start to change lanes ahead of the merge segment. In terms of behavior by group, except for basic segment (p<0.1) there was no significant difference in the number of lane changes between the two groups. (a) (b) Figure 6 3 Lateral Behavior by Freeway Segment 0.00 1.00 2.00 3.00 4.00 Basic Approach Merge Headway (mean) s Group A Group B 0.00 1.00 2.00 3.00 4.00 5.00 Basic Approach Merge SD (Longitudinal Acceleration) m/s 2 Group A Group B 0.00 0.50 1.00 1.50 2.00 2.50 3.00 Basic Approach Merge Lane Changes Group A Group B 0.55 0.60 0.65 0.70 0.75 0.80 Basic Approach Merge Lane Position Offset (m) Group A Group B

PAGE 80

80 Figure 6 from the Smaller values represent better lane position maintenance. There is significant drop (p<0.05) in overall lane position offset from basic to appro ach and merge segments. This can be attributed to increased situational awareness (Figure 5 1a). Hence it can be concluded from the results above that : 1 T he type of freeway section influence s the driver behavior 2 The Group B drivers consistently maintain lower headways than Group A drivers across all type s of freeway segments, while producing less longitudinal oscillations and better lane maintenance RQ3 Do Changes In Traffic Density Influence Driver Behavior And If Yes, How? Figure 6 4 and Figure 6 5 show longitudinal and lateral behavior respectively, of the two driver groups under different freeway segments. Longitudinal Behavior There are two differences to note from Figure 6 4 a: 1 There is a significant drop (p<0.05) in headways from low to medium density traffic. However, the difference between medium to high density is insignificant. 2 Headway s for group B drivers are significantly (p<0.05) less than that of group A drivers at all levels of traffic density This is consistent with the observation from control conditions (Figure 6 1) for CC1 (minimum safe headway during following).

PAGE 81

81 Similarly there are two differences to note from Figure 6 4 b : 1 There is no significant change in Standard Deviation of Longitudinal Acceleration (SDLA) between low to medium density. However, SDLA increases significantly (p<0.05) for both class es of drivers for High density traffic 2 Standard Deviation of Longitudinal Acceleration (SDLA) for group B drivers is significantly (p<0.05) less than that of group A drivers at all levels of traffic density This is consistent with the observation from control conditions for CC4 and CC5 (Longitudinal Oscillation expressed in terms of speed difference) (Figur e 6 1). Hence headway and SDLA measurements together indicate that the headways get shorter as the density increases and the longitudinal oscillation is highest at high density traffic. In all conditions, group B drivers (High SA/Low WL) maintain shorter headways with less oscillations i.e. better control. (a) (b) Figure 6 4 Longitudinal Behavior by Traffic Density Lateral Behavior Figure 6 5a gives the average number of lane changes by density. Group B drivers have statistically similar number of lane changes across different densities. However, for group A drivers the number of lane changes increases significantly (p<0.1) at high density condition One 0.00 1.00 2.00 3.00 4.00 Low Medium High Headway (mean) s Group A Group B 0.00 1.00 2.00 3.00 4.00 5.00 Low Medium High SD (Longitudinal Acceleration) m/s 2 Group A Group B

PAGE 82

82 possible reason for this is that group A drivers w ho experience high WL at high densities (Figure 5 8b) might prefer to stay on a certain lane to avoid heavier traffic Figure 6 There is no significant difference between Low and High density conditions between the groups of drivers. However, for the medium density condition Group B drivers maintain significantly (p<0.1) low er offset than low and high densities. This can be attributed to increased situational awareness (Figure 5 8 a). (a) (b) Figure 6 5 Lateral Behavior by Traffic Density Hence it can be concluded from the results above that: 1 The level of traffic density influences the driver behavior 2 Group B drivers consistently maintain lower headways than Group A drivers across at levels of traffic de nsities, while producing less longitudinal oscillations and better lane maintenance. RQ4 H ow Do Driving Regimes And The Minimum Safe Headway Vary By Driver Group And Traffic Condition ? Figure 6 6 shows the percentage of time that each of the two groups of drivers spent in four regimes ( free driving, following, approaching and speed compensation) over three different levels of density. The se regimes are defined as follows : 0.00 1.00 2.00 3.00 Low Medium High Lane Changes Group A Group B 0.00 0.20 0.40 0.60 0.80 Low Medium High Lane Position Offset (m) Group A Group B

PAGE 83

83 1. Free Driving (Blue in Figure 6 6) : This is when the car is either traveling at the desired speed already or accelerating to reach the desired speed, without any influence of a leader vehicle. 2 Following (Red in Figure 6 6) : This is when a leader vehicle is close to the subject vehicle (less than 5s headway), and the speed difference between the two vehicle s oscillates around zero 3 Approaching (Purple in Figure 6 6): T his is when the subject vehicle decelerates due to a slower leader vehicle . T change lanes. Sometimes, the leader vehicle could also change lanes, creating larger/open headways and a transition to free driving regime again. 4 Speed Compensation (Green in Figu re 6 6): If a vehicle is traveling below the desired speed, according to most car following models in the literature, the subject accelerates till the speed of the car matches with either the desired speed or that of the front vehicle speed (whichever is smaller) However, in our study it was observed in this data set that this is not always the case. S ome vehicles slowdown although the front vehicle is traveling faster than their speed. One explanation for this phenomenon is that the reduction in speed i s a mechanism to compensate for higher workload as discussed in Figure 5 3 Another explanation is that, the vehicle is decelerating in order to cr eate a sui gap in the neighboring lane I n this dataset i t wa s observed that such behavior is often followed by a lane change. These are the two observations to note from Figure 6 6: 1 As the vehicles have little choice but to follow a leader.

PAGE 84

84 (a) Group A (b) Group B Legend: Blue Free Driving, Red Following Purple Approaching/ Changing Lane Green Speed Compensation/Changing Lane Figure 6 6 Driving Regimes by Density 57% 6% 9% 28% Low Density 55% 11% 1% 33% Low Density 35% 22% 15% 28% Medium Density 29% 21% 2% 48% Medium Density 7% 54% 15% 24% High Density 15% 49% 3% 33% High Density

PAGE 85

85 2 Free Driving and Following regimes have similar proportion between the two groups of drivers. The difference between the two driver groups is in approaching and lane changing. The proportion of speed compensation regime is negligible in Group B, where a s it is significant in Group A. The Group B drivers approach a leading car, and either follow it or change lanes. Group A drivers may either choose the same options as Group B or might slow down before eventually changing lanes. While the distribution of d riving regimes gives an insight into combined (longitudinal and lateral) decision making of drivers, their quantitative impact on traffic is delivered through followin g a leader vehicle. This is the same as the CC1 parameter in VISSIM and has the most influence on the capacity of traffic flow. Figure 6 7a provides the minimum safe headway (CC1) at different levels of traffic for both driver groups. CC1 significantly dec reases for both driver groups as density increases. This is one of the observations that is not considered in most of the car following literature. According to most car following models, independent of the density of the surrounding traffic, drivers al ways maintain under various conditions. At all levels of densities, d river group B maintain significantly smaller (p<0.05) headways than driver group A. This follows the trend set in control scenario (Figure 6 1). The explanation for this is that group B drivers with higher SA and low WL are in a cognitively better state to control driving at sm aller headways.

PAGE 86

86 Figure 6 7b provides the minimum safe headway (CC1) at different segments of traffic for both driver groups. CC1 significantly decreases for both drivers groups at merge segments. This is the second observation that is not considered in most of the car following literature. According to most car following models, independent of the type of freeway segment, drivers always maintain the same minimum safe headway. This observation indicates tha t the minimum safe headway (and its inverse, maximum throughput) are affected by the type of segment. Based on these observations, drivers are willing to accept smaller headways around merge locations, which may in turn affect the lane changing behavior an d the respective probability of breakdown. At all types of segments, driver group B maintains significantly smaller (p<0.05) headways than driver group A. This follows the trend set in the control scenario (Figure 6 1). The explanation for this is that gr oup B drivers with higher SA and low WL are in a cognitively better state to control driving at smaller headways. (a) (b) Figure 6 7 Minimum Safe Headway by Level of Traffic and Type of Segment Research questions from the previous chapter have shown how the driver states i.e. WL and SA change across different freewa y se ctions under different levels of traffic density. Th is chapter has shown the same for driver behavior (CF and LC). The hypothesis from the Figure 3 0.00 0.50 1.00 1.50 2.00 2.50 Low Medium High Minimum Safe Headway (CC1) Group A Group B 0.00 0.50 1.00 1.50 2.00 2.50 Basic Approach Merge Minimum Safe Headway (CC1) Group A Group B

PAGE 87

87 1 is that changing conditions (traffic density type of segments) change driver cognitive states ( WL, SA) which in turn influence driver behavior ( CF LC ). Research Questions in this chapter ( RQ1 to RQ4 ) have demonstrated this th r ough use of two driver groups classified based on cognitive constructs (WL,SA) The next step is, if change in behavior is due to change in driver (cognitive) state, is it possible to model this relationship directly? The following section explores the answer to this question. RQ5 How To Model The Relationship Between Driver States And Driver Behavior? A nswers to the research questions ( RQ1 to RQ4 ) in this chapter demonstrate that changes in driver behavior (CF, LC) due to changing conditions consistently correspond to the changes in driver states (WL, SA). Moreover, the driver classes based on cognitive states show signifi cantly different behavior. If the effect of WL and SA on CF parameters can be directly modeled, it would be easier to incorporate relationship in microsimulation. For example currently in VISSIM minimum safe headway CC1 has a static value throughout the simulation irrespective of level of traffic density type of segment or type of drivers. If CC1 can be expressed in terms of W L and SA it could then be modeled as a dynamic variable dependent on prevailing conditions and driver state. Hence a n exploratory analysis was done to find a statistically significant relationship between car following parameters ( such as Safe minimum Headway CC1, Speed Oscillation CC4 and CC5 etc.) and driver states (WL, SA and combinations) A statistically significant relation ship was obtained between the terms CC1 Minimu m adjusted R 2 value of 0.46. The model is presented in Figure 6 8 Model details are presented next, followed by evaluation of CC1 values at boundary conditions to make sure that the model makes physical sense

PAGE 88

88 Model Figure 6 8 Relationship between CC1 and ( W L /SA) Figure 6 8 presents the relationship between WL/SA (Workload /Situational Awareness) and CC1 i.e. minimum safe headway for control (base) condition. Each point represents a driver coordinate and minimum safe headway as y coordinate. As shown, Group A drivers with low SA and high WL prefer to maintain larger headways than Group B drivers. The regression model has an adjusted R 2 value of 0.46. The relationship between Minimum Safe headway (CC1) and driver states (WL, SA) is given by : CC1 = 0.37+2.62 *(WL/SA) seconds ( Equation 6 1) Where, CC1 = Safe minimum headway in seconds WL = Workload ( Theoretical Range: 0% 100% Observed Range: 1 3 % 72% ) SA = Situational Awareness ( Theoretical Range: 0% 100% Observed Range: 26% 89% )

PAGE 89

89 Note: 1 It is practically impossible to experience 0% WL or have 100% SA 2 100% WL represents a condition where a driver is about to (or already) overwhelmed by the conditions. 0% SA presents a completely distracted or drowsy driver. Bounda ry Conditions And P hysical Meaning Although it would be ideal, i t is practically impossible to experience 0% WL or have 100% SA, and it is a concern for safety at 100% WL or 0% SA H ence conditions, let us evaluate what happens 1 Low WL: At WL close to 0 % CC1 is close to 0.37 seconds At very low workloads, the capability of the drivers is underutilized, hence they can afford to m aintain smaller headways. 2 High WL: At WL close to100 % CC1 is close to ( 0.37 +2.62=3) s econds (Assuming 100% SA) When workload gets ver y high the drivers seek to reduce it by maintaining larger headways. 3 High SA: At SA close to 10 0%, CC1 varies between to 0.37 to 3 seconds based on WL At very high levels of situational awareness drivers maintain headways based on WL. Lower WL leads to smaller headways and larger WL leads to larger headways 4 Low SA: At SA close to 0%, CC1 tends to a very large numbe r When drivers are distracted or drowsy, they need to maintain a large headway to account for slower reaction time. This can be achieved through either reducing the speed (compensation) or by changing to a lane with large r gap. E valuating the boundary conditions one can conclude that Equation 6 1 presents a practically sensible model. Hence, with more data, similar such models can be built.

PAGE 90

90 The trend line for Equation 6 1 in Figure 6 8 can also help differentiate between conservative and aggressive drivers safe and risky drivers relative to their driver states. For example: 1 Drivers above the trend line tend to maintain larger headways in relation to their WL/SA than an average driver from the data set. Hence they can be called relatively conservative 2 Drivers below the trend line tend to maintain smaller headways in relation to their WL/SA than an average driver from the data set. Hence they can be called relatively aggressive 3 Drivers below a certain safe headway threshold (for example say 1s) and to the extreme right of the figure represent risky drivers. These drivers have high WL/SA ratio which means they do not have the necessary cognitive capability to maintain smaller headways yet they pref er to drive with smaller headways. While equation 6 1 is helpful in simplifying the relationship between driver states and driver behavior, it assumes that all external factors that affect driver behavior are influenced by change in driver states. Further research is necessary in this regard to verify to what extent this assumption is true. Model D iscussion The following are the conclusions from this chapter : 1 Car following parameters estimated from control scenario tra jectories revealed that groups of drivers classified based on their driver states (WL, SA) indeed exhibit different driver behavior. 2 Mean headways and SD of Acceleration were used to study Longitudinal Behavior and it was revealed that headways get smaller with increase in density. SDLA was found to be

PAGE 91

91 maximum for high density conditions. Driver group B consistently maintained smaller headways and had better longitudinal control. 3 Numb er of Lane Changes and Lane Position Offset were used to study Lateral Behavior and it was revealed that most lane changes are made in approach segment and there is no significant difference between the driver groups in lane changing Using d river g roup B was considered to have better lateral control. 4 An exploratory analysis was conducted to develop a statistically significant direct relationship between Safe minimum Headway CC1 and driver states (WL, SA) The hypothesis from the Figure 3 1 is that changing conditions (traffic density, type of segments) change driver cognitive states (WL, SA) which in turn influence driver behavior (CF, LC). The exploration of the research questions from the pre vious chapter and this chapter have shown : 1 H ow the driver states i.e. WL and SA change across different freeway sections under different levels of traffic density. 2 H ow the driver behavior i.e. CF and LC change s across different freeway sections under different levels of traffic density. 3 How to model the relationship between driver states and driver behavior The next link in Figure 3 1 is the relationship between driver characteristic s and driver behavior. What driver characteristics make drivers a Group A driver or a Group B driver? Would a traditional Driver Behavior Questionnaire be helpful in differentiating these groups? The following chapter will focus on these questions.

PAGE 92

92 CHAPTER 7 DRIVER BEHAVIOR QUESTIONNAIRE (DBQ) SURVEY Collectin g driver information through questionnaires is one of the simpler methods to understand drivers and driver types. Two types of questionnaire were administered during the data collection: 1 Pre screening questionnaire before the driving simulator experiment asking for basic demographic information and driving preferences. 2 Traditional Driver Behavior Questionnaire (DBQ) after the driving simulator experiment n chester DBQ (Re ason et al, 1990) The aim of th is chapter is to utilize the available information and answer key questions such as, what driver characteristics make drivers a Group A driver or a Group B driver and would a traditional Driver Behavior Questionnaire be helpful in differentiating these groups? Based on the above discussion, the objectives for this chapte r are to : 1 Analyze pre screening data and identify differentiable properties between the driver groups 2 Analyze DBQ data and find the differences between two driver groups 3 Conduct Factor analysis and find the underlying core factors for a subset of DBQ questions Driver Group P roperties Figure 7 1 gives the p roperties of driver classes based on the pre screening questionnaire For ordinal variable s the Mann Whitney U test was used to check whether there were significant differences between Group A and Group B drivers. For scale variables, t test was used. Variables which are significant at 95% confidence level are indicated by double asterisks and those at significant at 90% are indicated through s ingle asterisk.

PAGE 93

93 Property Group A Group B Significan t? Comments N 18 22 WL (PDT) 49.2% 26.3% ** Measured SA 60.5% 73.1% ** Measured Age 42.7 36.8 Gender 39% 61% 59% 41% % Male % Female Frequency of Driving 2.2 2.5 1= <4 hr/week 4 = >14hr/week GPS usage 2.2 2 1= Always, 4 = Never Desired Freeway Speed 2.9 3.6 4 point scale 65 65 70,70 75,75+ Enjoy Driving 7.1 7.4 10 point scale Aggressive 4.1 5.4 10 point scale Tickets 5.50% 22.70% At least one in the past 1 year Involved in Accidents 33.30% 27.30% At least one in the past 5 years Significant at 90% CF, **Significant at 95% CF Figure 7 1 Properties of driver classes based on pre screening questionnaire Since the groups are classified based on WL and SA, the differences are indeed significant (p<0.05). In addition to these, group B drivers tend to be significantly younger (p<0.1) and more aggressive (p<0.1) than group A drivers. The desired speed selection on a 70 mph freeway lies between 70 to 75+ for Group B drivers, whereas group A drivers choose a lower desired speed (65 to 75). Although Group B drivers get significantly higher number of tickets (p<0.1), their accident record is statistically same as Group A drivers. Group B also drive more and enjoy driving, but the differences are statistically insignificant. The statistic on the number of tickets along with aggressiveness score indicates that those might have been aggressive violations such as speeding etc. The smaller headways (i.e.

PAGE 94

94 lar ger capacity) produced by Group B does not seem to come at a cost of safety, since they actually have fewer accident s than group A drivers. While this questionnaire is useful in obtaining a general driver profile, the next step is to get further insights into driving styles and how those affect driving behavior. DBQ surveys are traditionally designed to understand violation, errors and lapses by drivers T he objective of the next se ction is to explore if such DBQ survey would be helpful in differentiating the drivers who are classified based on cognitive states. DBQ R esponses Human factors in driving can be separated into two parts: driving skills and driving style, which both contribute to driving behavior. Driving skills (including the information processing and the maneuver executing) can be improved with the accumulation of driving experience. The driving style is the way a driver chooses to drive and is related to individua l driving habits (Li, 2014) To get an insight into individual driving style s a driver behavior questionnaire was administered at the end of the driving simulator experiment. The DBQ was first prepared by formulated their own questions based on this study to address their needs. The questions in these surveys are generally separated into three main c ategories: Violations, Errors, Slips and Lapses. While violations can be the result of social and motivational factors, errors, slips and lapses may be accounted by the information processing characteristics of the individual driver. Though the questions p repared for this study are based on the Ma n chester DBQ (Reason et al. 1990 ) and similar work such as Li (2014) etc. the intent here is to connect the responses for these questions with their measured driver states (WL and SA).

PAGE 95

95 Figure 7 2 DBQ Responses by Driver Groups The DBQ responses are presented in the following section, followed by factor analysis and discussion. Figure 7 2 shows the questions and responses by the driver groups. For every question, the participants were asked to estimate the level of likelihood on a five point scale (for

PAGE 96

96 frequency: 1=Never, 2= Nearly Never, 3= Seldom, 4= Sometimes, 5= Often; for likel ihood: <10%, 10%~40%, 40%~60%, 60%~90%, >90%). Since every item in Figure 7 1 is an ordinal variable, Mann Whitney U test wa s used to check if there were significant differences between Group A and Group B drivers in their DBQ responses. For Questions 4, 6, 7 and 13 there were statistically significant differences at 90% confidence level. However, apart from these four question s, there were no other statistically significant differences between Group A and Group B drivers in their DBQ responses. One possible reason for this is that the DBQ questions based on Manchester DBQ focus on driving styles, aggressiveness and errors wher eas the driver classification is based on driver states (i.e. WL and SA) experienced by the drivers. While an aggressive driver and a conservative driver might give significantly different responses to DBQ questions, they might be categorized into the same group if they experience similar WL and SA while driving. Though the differences in responses between the two groups are not statistically significant, conducting factor analysis with a subset of questions could help understand the discrepancy between gro up A and group B drivers. In the following section factor analysis results are presented along with discussion. Factor Analysis Exploratory factor analysis (EFA) is one of the most widely used statistical methods in psychological research, that goes beyond the individual items of tests and questionnaires to reveal the latent structure that underlies them. In the literature different researchers recommend 50 to 200 minimum data points for this type of analysis. However, by selecting a limited number o f correlated items a well conditioned dataset can yield reliable results (deWinter et al., 2009). or factors certain set of questi ons. While the factor analysis reveals optimum number of components and

PAGE 97

97 shows their statistical significance, the interpretation of these judgement. For this dataset, various subsets of the 27 Driver Behavior Questions were tested and ten questions were shor tlisted to achieve statistically significant components. The details of the EFA is as follows : Principal C omponents A nalysis ( PCA ) Principal Component Analysis (PCA) technique was used to find the components. The following are the statistical tests: Kaiser Meyer Olkin (KMO) Measure of Sampling Adequacy = 0.568. Normally, 0 < KMO < 1 If KMO > 0.5, the sample is adequate. Here, KMO = 0.568 which indicates that the sample is adequate and we may proceed with the Factor Analysis. Bartlett's Test of Spheric ity. Chi Square = 74.241 df = 45 p = 0. 004 The p value (Sig.) of 0.004 < 0.05, therefore the Factor Analysis is valid Eigen values (Select those components with Eigen Values >= 1) The initial components are the numbers of the variables used in the Factor Analysis. However, not all the 1 0 variables will be retained. The optimal number of component s should be at the turning point of the Scree plots and the Eigenvalue should be larger than 1. As shown in Figure 7 3 the optimum number of components is 3. Hence i n the present research only the 3 factors will be extracted by combining the relevant variables.

PAGE 98

98 Figure 7 3 Eigenvalue S cree P lot for DBQ data The Eigen values are the variances of the factors. The first factor will always account for the most variance and hence have the highest Eigen values. The next factor will account for as much of the left over variance as it can and the same will continue till the last factor. The percentage of variance represents the percent of total variance accounted by each factor and the cumulative percentage gives the cumulative percentage of variance account by the prese nt and the preceding factors. In the present research the first 3 factors explain 56.2 % of variance (Figure 7 4 ) The rotation sums of the squared loading represent the distribution of the variance after the varimax rotation with Kaiser Normalisation. The varimax rotation tries to maximize the variance of each of the factor.

PAGE 99

99 Component Initial Eigenvalues Extraction Sums of Squared Loadings Rotation Sums of Squared Loadings Total % of Variance Cumulative % Total % of Variance Cumulative % Total % of Variance Cumulative % 1 2.608 26.081 26.081 2.608 26.081 26.081 2.035 20.345 20.345 2 1.777 17.766 43.847 1.777 17.766 43.847 1.800 18.002 38.347 3 1.240 12.401 56.248 1.240 12.401 56.248 1.790 17.901 56.248 4 .967 9.670 65.918 5 .899 8.993 74.911 6 .705 7.048 81.959 7 .677 6.773 88.733 8 .472 4.723 93.456 9 .369 3.689 97.144 10 .286 2.856 100.000 Extraction Method: Principal Component Analysis. Figure 7 4 Total Var iance Explained by Each of the C omponents The variables included in each factor and their factor loading are show n in the rotated component matrix in Figure 7 5 The factor loading above 0.4 are ignored. The Rotated Factor Matrix represents the rotated factor loadings, which are the correlations between the variables and the factors. The next step is to identify the qualitative meaning of the three factors. Compon ents 1 2 3 Q 1 .568 Q 3 .794 Q 4 .656 Q 7 .540 Q 10 .634 Q 11 .783 Q 15 .629 Q 20 .631 Q 23 .755 Q 27 .793 Figure 7 5 Rotated Component Matri x with Factor Loadings

PAGE 100

100 Based on the questions and their qualitative meaning, the three core factors from the factor analysis can be described as: 1. Situational Awareness/Distraction (Questions 1, 3, 4 and 7) 2. Workload Adjustment/Compensation (Questions 20, 23 and 27) 3. Errors and Misjudgments (Questions 10, 11, 15) Factor 1 Situational Awareness/D istraction Figure 7 6 Factor 1 Responses Questions ( 1=Never, 2= Nearly Never, 3= Seldom, 4= Sometimes, 5= Often ) 1. Do you check your speedometer and discover that you are unknowingly speeding? 3. Do you attempt to drive away without having switched on the ignition? 4. Do you get distracted and realize belatedly that the vehicle ahead is slow and slam the break to avoid collision? 7. Do you try to overtake without checking the mirror and get hoote d by the car behind? Figure 7 6 gives DBQ responses for Factor 1 questions. The higher numbers represent less awareness or more distraction. Except for Question 3, all other questions deal with 3.94 1.22 2.56 1.83 3.64 1.36 2.05 1.50 1 3 4 7 Factor 1 Situational Awareness/Distraction Group A Group B

PAGE 101

101 awareness during driving (i.e once the car has been started) The difference s in responses to Q4 and Q7 were found to be statistically significant (p<0.1). Fr om questions 1, 4 and 7 it is c lear that Group A drivers state that they are less aware during driving than Group B drivers This is consistent with their act ual SA measurements, where Group A were found to be less aware than Group B. Factor 2 Workload Adjustment/ C ompensation Figure 7 7 Factor 2 Responses Questions ( 1=Never, 2= Nearly Never, 3= Seldom, 4= Sometimes, 5= Often ) 20. When you are followed by or are following a big truck, do you feel unsafe and try to change lanes as soon as possible? 23. In congestion, do you lose patience to wait for a sufficient gap and make a forced lane change to the target lane? 27. Do you keep driving on the left most l ane for long time? Figure 7 7 gives DBQ responses for Factor 2 questions. The higher numbers represent drivers experiencing higher workload in certain situations and deci ding on a compensatory action (e xample: s hifting to a target lane) Though Question 27 statistically correlates to factor 2, physically moving to left can be either to decrease WL or to gain speed advantage. Hence Q27 3.61 2.72 3.44 3.09 2.59 3.45 20 23 27 Factor 2 Workload Adjustment/Compensation Group A Group B

PAGE 102

102 responses could not be necessarily considered as compensation for increased workload. However, f rom questi ons 20 and 23 it is clear that Group A drivers state t hat take compensatory actions while driving during increased workloads (i.e. congestion, trucks etc.) than Group B drivers. This is consistent with their actual WL and Speed measurements where Group A were found to be decreasing their speeds at higher traffic densities compared to Group B. Factor 3 Errors And Misjudgments Figure 7 8 Factor 3 Responses Questions ( 1=Never, 2= Nearly Never, 3= Seldom, 4= Sometimes, 5= Often ) 10. Do you misjudge speed of surrounding vehicles while making lane changing maneuvers? 11. Do you fail to read or misread the signs while driving? 15. Do you misjudge the gaps available to make driving maneuvers? Figure 7 8 gives DBQ responses for Factor 3 questions. The hi gher numbers represent more errors From questions 1 0 11 and 15 it is clear that Group A drivers state that they are more likely to make certain errors during driving than Group B drivers. This is consistent with their actual WL and SA measurements, where Group B were found to have more SA and 2.00 2.39 1.94 1.77 2.18 1.77 10 11 15 Factor 3 Errors and Misjudgments Group A Group B

PAGE 103

103 experience less WL than Group A which are the ideal conditions for driving safely and efficiently Factor Analysis D iscussion The following are the conclusions from this chapter : 1 After the driving simulator experiment, the participants were asked to fill out a DBQ survey designed on the basis of Manchester DBQ (Reason et al., 1990) 2 Mann Whitney test revealed significant differences between Group A and Group B drivers for only four questions (Questions 4, 6, 7 and 13) at 90% confidence leve l. No significant differences were found between the two driver groups for the rest of the DBQ responses This can be attributed to differences in the nature of DBQ surveys and driver cla ssification. 3 An exploratory factor analysis revealed three core factors from ten questions. They are: Situation Awareness/Distraction Workload Adjustment/Compensation, Errors and Misjudgments. For each of the factors, the differences in responses fo r the two driver groups were found to be consistent with their driver states (WL and SA) 4 In the future, d esigning a disparate DBQ aimed at revealing drivers actions regarding their workload and situational awareness could be useful in revealing statistically significant differences between driver groups.

PAGE 104

104 CHAPTER 8 CONCLUSIONS AND FUTURE WORK Figure 8 1 presents a framework relating driver characteristics and performance to traffic operations and highway design. This framework was used as the basis for this dissertation. The lower part of the figure presents microscopic elements of the system, while the top presents macroscopic elements. The solid lines represent the models and relationships that are established in the field of traffic operations. The dashed lines represent the models and relationships being studied in the field of psychology; their usage is virtually non existe nt in traffic operational analyses (until now) Figure 8 1 Elements of Traffic Operations (Reproduced from Figure 3 1) The objective of this dissertation was to consider these relationships (dashed lines from Figure 8 1) that were theorized in the field of psychology and to examine them using driver

PAGE 105

105 simulator data and establish a framework to incorporate human behavior elements in traffic microsimulation Summary 1 Li terature from Psychology and Traff ic Engineering were studied and based on that States 2 Based on the literature in psychology and human factors, a driving simulator experiment was designed to collect the driver measures at three level of traffic densities (Low, Medium, High) and across three types of freeway segments (Basic, A pproach, M erge). Driver Behavior Questionnaire was ad ministered and trajectory data was saved. 3 A novel estimation methodology was developed to directly measure car following parameters from a controlled following trajectory 4 The differences in driver states (WL, SA) between each freeway segment as well as level of traffic density were found to be statistically significant. 5 Cluster analysis was used to classify drivers into two groups based on SA and WL. The first group of drivers had significantly lower SA and experienced higher WL than the second group. The WL and SA patterns for the two driver groups were found to be consistent across types of freeway segments and levels of traffic density. 6 C F parameters estimated from control scenario trajectories revealed that groups of drivers classified based on their driver states (WL, SA) indeed exhibit different driver behavior. 7 An exploratory analysis was done to find a statistically significant direct relationship between Safe minimum Headway CC1 and driver states (WL, SA). 8 Exploratory Factor Analysis was done to reveal the core factors from the DBQ questions.

PAGE 106

106 Conclusions The following are the main conclusions from this dissertation: 1 Changes in traffic density and type of freeway segment affect a driver It was found that medium density traffic produces ideal combination of WL and SA (i.e High SA, relatively low WL). It was also found that drivers use compensatory mechanism such as reducing speed to negotiate higher levels of WL. 2 Drivers can be classified based on their WL and SA l e vels. Driver from such classification show distinct ly different driver behavior. The two drive r groups were compared with their longitudinal and lateral behavior. Driver group B who have higher SA and experience lower WL consistently maintained smaller headways and had better longitudinal and lateral control. 3 It is possible to directly model the relationship between driver states and driver behavior. Minimum Safe H eadway was modeled as a function of WL and SA. 4 An exploratory factor analysis revealed three core factors from ten DBQ questions. Except for four questions no significant differences were found between the two driver groups for the rest of the DBQ responses. This can be attributed to differences in the nature of DBQ surveys and driv er classification. However, a number of driver properties (age, aggressiveness sco re, speed selection etc.) were significantly different for the two groups. Future Work The f ollowing are potential paths for future work: 1 Collecting and analyzing more trajectory data to build models similar to the one presented in Chapter 6 with an eventual objective of i mplementing either a model or a set of parameters based on the collected data in a microsimulation package :

PAGE 107

107 F rom Chapter s 5 and 6 we know that turn influence driver behavior (CF and LC). Some microsimulation packages collect information of surrounding vehicles and this could be used to model influence of real time. Anot her example application would be we know from chapter 5 and 6 that type of segment influences WL, SA and in turn CF and LC. Some microsimulation packages allow users to define driver behavior by segment or a link ters can be implemented separately by segments. 2 Using the existing data to test the proposed models in the literature such as Hoogendoorn et al (2013), S aifuzzaman and Zheng (2014) etc : Both of these papers propose modification to existing car following models through a set of terms to represent driver states. WL and SA are addressed in these papers but with slightly different terminology. The limitations of these paper has been that the proposed relationships have not been validated wi th any data (field or simulated ). The data from driving simulator experiment could be selected to match the definitions of driver states components in these papers and these models can be validated. Based on the results of the validation, our own CF model modifications can be proposed. 3 Creating driver classes based on DBQ data and comparing them with the driver groups based on WL and SA and developing a non traditional Driver Behavior Questionnaire focused on differentiating drivers based on their WL and SA behavior etc. Driving simulator experiments can be expensive and time consuming to consider the effects of behavior DBQ survey presents a more viable alternative where the driver states and driver groups could be deduced by asking a specific set of questions. Chapter 7 revealed that traditional Manchester DBQ (Reason et al,

PAGE 108

108 1990) type of a survey was not very helpful in differentiating driver state (WL and SA) based driver groups Hence a set of DBQ survey questions focusing on how aware drivers are how much workload they experience and what (compensatory) actions they take to negotiate commonly occurring traffic situations could be much more useful in differentiating a high WL, low SA driver group from a low WL, high SA driver group. For example, questions Q4 and Q7 revealed significant differences between the two groups of drivers and the factor analysis revealed the underlying core factor being situational awareness. The future work is to develop more such questions. To conclude, this dissertation validates some of the proposed theories and relationships in psychology as applicable to driving behavior a nd presents a framework to begin bridging the gap between two connected fields (psychology and traffic operations) that have grown apart.

PAGE 109

109 APPENDIX A PARTICIPANT R ECRUITMENT F LYER

PAGE 110

110 APPENDIX B PRESCREEN Q UESTIONNAIRE To Participants: This questionnaire is used to select a diverse pool of drivers to participate in the driving simulator experiment for studying driver behaviors along freeways. Information collected in this form will be used for traffic engineering research only. All respo nses will be held in confidence and are exempted from public disclosure 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 respon ses will not be disclosed in identifiable form without your consent. Please respond to as many of the following questions as possible. 1. What is your gender? Male Other Female 2. What is your age range? < 20 20 30 30 40 40 50 50 60 >=60 3. Which of the following groups do you most identify yourself as? Caucasian Native American African American Pacific Islander Hispanic Asian Other or prefer not to answer _______ 4. Where did you begin your driving practice and obtained your driver license? North America Latin America Asia Europe Australia Other __________ 5. How long have you been driving in the U.S.? < 1 year 3 to 9 years 1 to 3 years >=10 years 6. Do you have a valid U.S. driver license? Yes No

PAGE 111

111 7. What is your occupation? Full time student Professional driver University faculty/staff Other, explain___________ 8. How often do you drive to work/school? Everyday Sometimes Usually Never 9. How often do you use GPS device/ phone navigation? Everyday Sometimes Usually Never 10. How much time do you spend driving per week? < 4 hours 8 to 14 hours 4 to 8 hours > 14 hours 11. What type of vehicle do you usu ally drive? Sedan/Coupe Jeep Pickup/SUV Truck 12. When the speed limit on a freeway is 70 mph, what speed are you likely to drive at (assuming good visibility and good weather conditions) ? less than 65 mph 65 to 70 mph 70 to 75 mph more than 75 mph 13. How often do you change lanes to gain speed advantage? whenever possible

PAGE 112

112 often seldom never 14. On a scale of 1 to 10, how much you enjoy driving? 6 15. On a scale of 1 to 10, how aggressive you rate yourself as a driver? 2 16. On a scale of 1 to 10, how aggressive your family/friends rates you as a driver? 2 17. Did you get a traffic ticket of any kind in the last year? Y/N if yes describe 0 18. Were you involved in an accident in the last five years? Y/N if yes describe 0 19. n (at least one from phone/email) Name: ________________ Phone: ________________ Email: ________________ Date: ________________ Return Address: By Email: pruthvim@ufl.edu

PAGE 113

113 APPENDIX C SIMULATOR S ICKNESS Q UESTIONNAIRE (SSQ) Circle the item that best represents how you are feeling regarding each symptom: General discomfort None Slight Moderate Severe Fa tigue None Slight Moderate Severe Headache None Slight Moderate Severe Eyestrain None Slight Moderate Severe Difficulty focusing None Slight Moderate Severe Increased salivation None Slight Moderate Severe Sweating None Slight Moderate Severe Nausea None Slight Moderate Severe Difficulty concentrating None Slight Moderate Severe Fullness of head None Slight Moderate Severe Blurred vision None Slight Moderate Severe Dizzy (eyes open) None S light Moderate Severe

PAGE 114

114 APPENDIX D NASA TLX QUESTIONS

PAGE 115

115

PAGE 116

116 APPENDIX E SITUATIONAL AWARENESS GLOBAL ASSESSMENT TEST (SAGAT) QUESTIONS 1.1 What is the Speed Limit? 1.2 Which lane is nearest vehicle traveling in? 1.3 How will the distance to the vehicle in front (in the same lane) change in the next few seconds? 2.1 Were you traveling above or below the speed limit? 2.2 What is the relative speed of the traffic compared to our vehicle? 2.3 Are you losing or gaining on the vehicle in front (in the same lane)? 3.1 Which lane are you currently in? 3.2 If you maintain current acceleration, are you going to drive below or above the current speed limit? 3.3 What is the speed of the vehicle in front (in the same lane)? 4.1 Are there any vehicles in the ri ght/left lane adjacent to your car? 4.2 Are the brake lights of lead (vehicle in front) vehicle on? 4.3 How far have you driven till now (approximate to the nearest 0.5 mile)? 5.1 How many lanes are occupied in front? 5.2 What was the last road sign you sa w? (Speed limit, route signs, etc) 5.3 How long have you driven till now (approximate to the nearest 0.5 minute)? 6.1 What speed is your vehicle traveling? 6.2 How many times have you changed lanes (since starting at ramp on this drive)? 6.3 How far till t he end of drive was left (approximate to the nearest 0.5 minute)? 7.1 Have speed limits changed in this drive? 7.2 Which lane has the most gap in the front? 7.3 If you maintain your current speed, is it possible that we will have a collision with lead vehicle? 8.1 Is speed of the closest vehicle in front of above/below the current speed limit? 8.2 What is the distance between your car and the vehicle in front (in the same lane)? (Hint: 1 car length = 20 feet) 8.3 When will you reach the nearest traffic sign (approximate to the nearest 0.5 minute)? 9.

PAGE 117

117 APPENDIX F DRIVER B EHAVIOR Q UESTIONNAIRE (DBQ) For the questions below estimate the level of likeli hood on a five point scale (for frequency: 1=Never, 2= Nearly Never, 3= Seldom, 4= Sometimes, 5= Often; for likelihood: <10%, 10%~40%, 40%~60%, 60%~90%, >90%). 1. Do you check your speedometer and discover that you are unknowingly speeding? 2. Do you become impatient with a driver in the outer lane and overtake on the inside? 3. Do you attempt to drive away without having switched on the ignition? 4. Do you get distracted and realize belatedly that the vehicle ahead is slow and slam the break to avoid collision? 5. riving with no clear recollection of the road you just travelled? 6. Do you overtake in risky circumstances to avoid getting stuck behind a slower vehicle? 7. Do you try to overtake without checking the mirror and get hooted by the car behind? 8. Do you deliberatel y disregard the speed limit at early hours in the morning and late in the night? 9. Do you have an aversion to particular type of drivers? 10. Do you misjudge speed of surrounding vehicles while making lane changing manoeuvres? 11. Do you fail to read or misread the signs while driving? 12. 13. 14. 15. Do you misjudge the gaps available to make driving manoeuvres? 16. Do you sound horn to indicate your annoyance to other drivers? 17. Do you stay in a free way lane that you know will be closed ahead until the last minute before forcing your way into the other lane ? 18. D o you d rive so close to the car in front that it would be difficult to stop in an emergency ? 19. Do you a ttempt to overtake someone that you had not noticed to be signalling a right turn ? 20. When you are followed by or are following a big truck, do you feel unsafe and try to c hange lanes as soon a s possible? 21. Do you turn on the indication light when you drive off or change lanes? 22. Do y ou still conduct overtaking even when you see the preceding vehicle already turns on the left turning indicator? 23. I n congestion, do you lose patience to wait for a sufficient gap and make a forced lane change to the target lane ? 24. Do you yield for the driver whose lane will be end/closed ? 25. Do you yield for other drivers forced lane changing in congestion ? 26. Do you o ften check the mirrors when going straight on ? 27. Do you k ee p driving on the left most lane for long time ?

PAGE 118

118 APPENDIX G PERIPHERAL D ETECTION T EST (PDT) The figure shows the relationship between Workload and responses and reaction time of responses to the task. If the drivers fail to respond to the peripheral visual cue within 3 seconds f they respond correctly between 2 to 3 seconds, the miss rate increases from 0 to 100%. The Workload (PDT) is calculated as % Miss Rate. This was done for two rea sons: 1. To avoid overlap of successive visual cues, a reaction time had to be chosen to establish as a cut 2. To avoid giving equal weight to all the hits, a quicker reaction time had to be chosen for (2s) Figure: Relationship between reaction time and Workload (PDT) For each density, drivers encounter 12 14 such visual cues placed throughout the simulation at random intervals. The responses are recorded, separated and averaged based on the traffic d ensity and geometric segment. 0% 20% 40% 60% 80% 100% 120% 0 1 2 3 4 Reaction time (s) Workload PDT/ Miss Rate (%)

PAGE 119

119 LIST OF REFERENCES Aghabayk, K., Sarvi, M., Young, W., & Kautzsch, L. (2013). A novel methodology for evolutionary calibration of Vissim by multi threading. In Australasian Transport Research Forum (pp. 1 15). Ahn, S., Cassidy, M. J., & Laval, J. (2004). Verification of a simplified car following theory. Transportation Research Part B: Methodological 38 (5), 431 440. Arin, C., Jongen, E. M., Brijs, K., Brijs, T., Daniels, S., & Wets, G. (2013). A simulator study on the impact of traffic calming measures in urban areas on driving behavior and workload. Accident Analysis & Prevention 61, 43 53. Auberlet, J.M., A. Bhask ar, B. Ciuffo, H. Farah, R. Hoogendoorn, A. Leonhardt. (2014). Data collection techniques. Traffic Simulation and Data: Validation Methods and Applications Baldauf, D., Burgard, E., & Wittmann, M. (2009). Time perception as a workload measure in simulated car driving. Applied ergonomics 40(5), 929 935. Bando, M., Hasebe, K., Nakanishi, K., Nakayama, A., Shibata, A., & Sugiyama, Y. (1995). Phenomenological study of dynamical model of traffic flow. Journal de Physique I 5 (11), 1389 1399. Bando, M., Hasebe, K., Nakanishi, K., & Nakayama, A. (1998). Analysis of optimal velocity model with explicit delay. Physical Review E 58 (5), 5429. Barcel, J. (2010). Fundamentals of traffic simulation (Vol. 145, p. 439). New York: Springer. Baumann, M. R., Rsler, D., & Krems, J. F. (2007). Situation awareness and secondary task performance while driving. In International Conference on Engineering Psychology and Cognitive Ergonomics (pp. 256 263). Springer Berlin Heidelberg. Bekiaris, E., Amditis, A., & Panou, M. (2003). DRIVABILITY: a new concept for modelling driving performance. Cognition, Technology & Work 5 (2), 152 161. Bella, F. (2008). Driving simulator for speed research on two lane rural roads. Accident Analysis & Prevention 40(3), 1078 1087. Ben Akiva, M., Davo l, A., Toledo, T., Koutsopoulos, H. N., Burghout, W., Andrasson, I., & Lundin, C. (2002, January). Calibration and evaluation of MITSIMLab in Stockholm. In 81st Transportation Research Board Meeting Blana, E. (1996). Driving simulator validation studies: A literature review. Institute of Transport Studies, University of Leeds, Working Paper 480 Transportation Research Part F: Traffic Psychology and Behaviour 2 (4), 201 206.

PAGE 120

120 Bolstad, C. A. (2 001). Situation awareness: does it change with age?. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting (Vol. 45, No. 4, pp. 272 276). Brackstone, M., & McDonald, M. (1999). Car following: a historical review. Transportation Research Part F: Traffic Psychology and Behaviour, 2 (4), 181 196. Cantin, V., Lavallire, M., Simoneau, M., & Teasdale, N. (2009). Mental workload when driving in a simulator: Effects of age and driving complexity. Accident Analysis & Prevention, 41(4), 763 771. Carsten, O. (2007). From driver models to modelling the driver: what do we really need to know about the driver?. In Modelling driver behaviour in automotive environments (pp. 105 120). Springer London. Chang, M., Kim, J., Kang, K., & Yun, J. (2001). Evalu at freeway merging area. In Proceedings of the 80th Annual Meeting of the Transportation Research Board De Waard, D. (1996). The measurement of drivers' mental workload Netherlands: Groningen University, Traffic Research Center. De Winter, J., Van Leeuwen, P. M., & Happee, R. (2012). Advantages and disadvantages of driving simulators: a discussion. In Proceedings of Measuring Behavior (pp. 47 50). Dingus, T. A., Klauer, S. G., Neale, V. L., Petersen, A., Lee, S. E., Sudweeks, J. D., & Bucher, C. (2006). The 100 car naturalistic driving study, Phase II results of the 100 car field experiment (No. HS 810 593). following model using mixed ve hicle pair interactions. Transportation research part C: emerging technologies 67 227 242. Endsley, M. R. (1993). Situation awareness and workload Flip sides of the same coin. In International Symposium on Aviation Psychology, 7 th, Columbus, OH (pp. 90 6 911). Endsley, M. R. (1995). Toward a theory of situation awareness in dynamic systems. Human factors, 37(1), 32 64. Endsley, M. R. (2017). From here to autonomy: lessons learned from human automation research. Human factors, 59(1), 5 27. Filtness, A. J. & Rudin Brown, C. M. (2012). Drinking and riding: is subjective workload related to performance? In Australasian Road Safety Research Policing Education Conference, Wellington, New Zealand. Fritzsche, H. T. (1994). A model for traffic simulation. Traffic Engineering+ Control 35 (5), 317 21.

PAGE 121

121 Fuller, R. (2005). Towards a general theory of driver behaviour. Accident Analysis & Prevention 37 (3), 461 472. Gazis, D. C., Herman, R., & Rothery, R. W. (1961). Nonlinear follow the leader models of traffic flow. Op erations research 9(4), 545 567. Gipps, P. G. (1981). A behavioral car following model for computer simulation. Transportation Research Part B: Methodological 15 (2), 105 111. Godley, S. T., Triggs, T. J., & Fildes, B. N. (2002). Driving simulator validation for speed research. Accident analysis & prevention 34(5), 589 600. Hamdar, S. (2012). Driver behavior modeling. In Handbook of Intelligent Vehicles (pp. 537 558). Springer London. Hancock, P. A., Wulf, G., Thom, D., & Fassnacht, P. (1990). Driver workload during differing driving maneuvers. Accident Analysis & Prevention, 22(3), 281 290. Hancock, P. A., Lesch, M., & Simmons, L. (2003). The distraction effects of phone use du ring a crucial driving maneuver. Accident Analysis & Prevention, 35(4), 501 514. Hart, S. G., & Staveland, L. E. (1988). Development of NASA TLX (Task Load Index): Results of empirical and theoretical research. Advances in psychology, 52, 139 183. Hidas, P (2005). A functional evaluation of the AIMSUN, PARAMICS and VISSIM microsimulation models. Road and Transport Research 14 (4). Higgs, B., Abbas, M., & Medina, A. (2011). Analysis of the Wiedemann car following model over different speeds using naturalist ic data. Road Safety and Simulation, Indianapolis, Indiana Hoogendoorn, R., van Arem, B., & Hoogendoorn, S. (2013). Incorporating driver distraction in car following models: Applying the TCI to the IDM. In Intelligent Transportation Systems (ITSC), 2013 1 6th International IEEE Conference on (pp. 2274 2279). IEEE. Hoogendoorn, R., B. Van Arem. (2013) "Modeling traffic system complexity through fuzzy entropy." In 16th International IEEE Conference on Intelligent Transportation Systems 540 546. Jahn, G., Oehm e, A., Krems, J. F., & Gelau, C. (2005). Peripheral detection as a workload measure in driving: Effects of traffic complexity and route guidance system use in a driving study. Transportation Research Part F: Traffic Psychology and Behaviour 8 (3), 255 275. Jansen, R. J., Sawyer, B. D., van Egmond, R., de Ridder, H., & Hancock, P. A. (2016). Hysteresis in mental workload and task performance: the influence of demand transitions and task prioritization. Human factors, 58(8), 1143 1157. Jiang, R., Wu, Q., & Zh u, Z. (2001). Full velocity difference model for a car following theory. Physical Review E, 64(1), 017101.

PAGE 122

122 Kass, S. J., Cole, K. S., & Stanny, C. J. (2007). Effects of distraction and experience on situation awareness and simulated driving. Transportation Research Part F: Traffic Psychology and Behavior, 10(4), 321 329. Kennedy, R. S., Lane, N. E., Berbaum, K. S., & Lilienthal, M. G. (1993). Simulator sickness questionnaire: An enhanced method for quantifying simulator sickness. The international journal of aviation psychology 3 (3), 203 220. Kondyli, A., & Elefteriadou, L. (2011). Modeling driver behavior at freeway ramp merges. Transportation Research Record: Journal of the Transportation Research Board (2249), 29 37. Kondyli, A., Elefteriadou, L., Brilon W., Hall, F. L., Persaud, B., & Washburn, S. (2013). Development and evaluation of methods for constructing breakdown probability models. Journal of Transportation Engineering, 139(9), 931 940. Lenz, H., Wagner, C. K., & Sollacher, R. (1999). Multi antic ipative car following model. The European Physical Journal B Condensed Matter and Complex Systems 7 (2), 331 335. Li, Jie. (2014). Characteristics of Chinese Driver Behavior TU Delft, Delft University of Technology. Lighthill, M. J., & Whitham, G. B. (195 5). On kinematic waves. II. A theory of traffic flow on long crowded roads. In Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences (Vol. 229, No. 1178, pp. 317 345). The Royal Society. Ma, R., & Kaber, D. B. (2005). Situation awareness and workload in driving while using adaptive cruise control and a cell phone. International Journal of Industrial Ergonomics 35 (10), 939 953. Manjunatha, P., Vortisch, P., & Mathew, T. V. (2013). Methodology for the Calibration of VISSIM in Mixed Traffic. In 92nd Annual Meeting of the Transportation Research Board. Washington, DC Manjunatha, P., Kondyli, A., & Elefteriadou, L. (2017). How Has Driver Behavior Been Considered in Traffic Microsimulation and How Can we Use Cognitiv e Sciences and Psychology Studies to Enhance Them? In 96th Annual Meeting of the Transportation Research Board. Washington, DC Manjunatha, P. & Elefteriadou, L. (2018). Analysis of Wiedemann Car Following Thresholds Using Driving Simulator Observations. I n 97th Annual Meeting of the Transportation Research Board. Washington, DC. Martin, J. L. (2002). Relationship between crash rate and hourly traffic flow on interurban motorways. Accident Analysis & Prevention, 34(5), 619 629.

PAGE 123

123 Menneni, S., Sun, C., & Vorti sch, P. (2008). Microsimulation calibration using speed flow relationships. Transportation Research Record: Journal of the Transportation Research Board (2088), 1 9. Michon, J. A. (1985). A critical view of driver behavior models: What do we know, what sh ould we do. Human behavior and traffic safety 485 520. Moridpour, S., Sarvi, M., & Rose, G. (2010). Lane changing models: a critical review. Transportation letters 2 (3), 157 173. Newell, A., & Simon, H. A. (1972). Human problem solving (Vol. 104, No. 9). Englewood Cliffs, NJ: Prentice Hall. Panwai, S., & Dia, H. (2005). Comparative evaluation of microscopic car following behavior. IEEE Transactions on Intelligent Transportation Systems 6 (3), 314 325. Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2008). Situation awareness, mental workload, and trust in automation: Viable, empirically supported cognitive engineering constructs. Journal of Cognitive Engineering and Decision Making, 2(2), 140 160. Patten, C. J., Kircher, A., stlund, J., Nilsson, L. & Svenson, O. (2006). Driver experience and cognitive workload in different traffic environments. Accident Analysis & Prevention, 38(5), 887 894. Pipes, L. A. (1953). An operational analysis of traffic dynamics. Journal of applied physics 24 (3), 274 281 Punzo, V., & Simonelli, F. (2005). Analysis and comparison of microscopic traffic flow models with real traffic microscopic data. Transportation Research Record: Journal of the Transportation Research Board (1934), 53 63. Punzo, V., Ciuffo, B., & Montan ino, M. (2012). Can results of car following model calibration based on trajectory data be trusted?. Transportation Research Record: Journal of the Transportation Research Board (2315), 11 24. PTV (2014), www.ptvgroup. com (Accessed 05/15/2016) Quadstone (2004). www.paramics online.com (Accessed 05/15/2016) Ranjitkar, P., Nakatsuji, T., & Kawamura, A. (2005). Experimental analysis of car following dynamics and traffic stability. Transportation Research Record: Journal of the Transportation Research Board (1934), 22 32. Reason, J., Manstead, A., Stradling, S., Baxter, J., & Campbell, K. (1990). Errors and violations on the roads: a real distinction. Ergonomics 33 (10 11 ), 1315 1332.

PAGE 124

124 Reimer, B., Mehler, B., Coughlin, J. F., Roy, N., & Dusek, J. A. (2011). The impact of a naturalistic hands free cellular phone task on heart rate and simulated driving performance in two age groups. Transportation research part F: traffic ps ychology and behavior, 14(1), 13 25. Saifuzzaman, M., & Zheng, Z. (2014). Incorporating human factors in car following models: a review of recent developments and research needs. Transportation research part C: emerging technologies 48 379 403. Salvucci, D. D. (2006). Modeling driver behavior in a cognitive architecture. Human factors 48 (2), 362 380. Shechtman, O., Classen, S., Awadzi, K., & Mann, W. (2009). Comparison of driving errors between on the road and simulated driving assessment: a validation s tudy. Traffic injury prevention, 10(4), 379 385. Soria, I., L. Elefteriadou, A. Kondyli. "Assessment of car following models by driver type and under different traffic, weather conditions using data from an instrumented vehicle." Simulation modelling practice and theory 40 (2014): 208 220. Srinivasan, S., Rogers, J., Manjunatha, P., Maldonado, P., & Yarney, A. (2017). Development of Support Systems, Instructional Modules, and a Case Study for the Enhanced Driving Simulator at the Gator Tech Smart House STRIDE Report, university of Florida. Sun, D., & Elefteriadou, L. (2012). A driver behavior based lane changing model for urban arterial streets. Transportation science 48 (2), 184 205. Treiber, M., Hennecke, A., & Helbing, D. (2000). Congested traffic s tates in empirical observations and microscopic simulations. Physical review E 62 (2), 1805. TSIS (2011), www.ce.mctrans.ufl.edu (Accessed 05/15/2016) TSS, A. (2009). 5.0 Microsimulator User's Manual. Transportation Simulation Systems Van Lint, H., Schake l, W., Tamminga, G., Knoppers, P., & Verbraeck, A. (2016). Getting the Human Factor into Traffic Flow Models: New Open Source Design to Simulate Next Generation of Traffic Operations. Transportation Research Record: Journal of the Transportation Research B oard (2561), 25 33. Verwey, W. B. (2000). On line driver workload estimation. Effects of road situation and age on secondary task measures. Ergonomics, 43(2), 187 209. Vidulich, M. A. (2000). Testing the sensitivity of situation awareness metrics in inter face evaluations. Situation awareness analysis and measurement, 227 246. Vidulich, M. A., & Tsang, P. S. (2012). Mental workload and situation awareness. Handbook of human factors and ergonomics, 243 273.

PAGE 125

125 Wickens, C. D. (2008). Situation awareness: Review of Mica Endsley's 1995 articles on situation awareness theory and measurement. Human factors, 50(3), 397 403. Wiedemann, R. (1974). Simulation des Strassenverkehrsflusses. Schriftenreihe des Instituts fr Verkehrswesen der Universitt Karlsruhe. Wilde, G. J. (1982). The theory of risk homeostasis: implications for safety and health. Risk analysis 2 (4), 209 225. Yan, X., Abdel Aty, M., Radwan, E., Wang, X., & Chilakapati, P. (2008). Validating a driving simulator using surrogate safety measures. Accident An alysis & Prevention, 40(1), 274 288. Yang, Q., & Koutsopoulos, H. N. (1996). A microscopic traffic simulator for evaluation of dynamic traffic management systems. Transportation Research Part C: Emerging Technologies, 4(3) 113 129. Zhao, X. A. G. Z., & Ga o, Z. Y. (2005). A new car following model: full velocity and acceleration difference model. The European Physical Journal B Condensed Matter and Complex Systems 47 (1), 145 150.

PAGE 126

126 BIOGRAPHICAL SKETCH Pruthvi Manjunatha received his Master of Technology degree from Indian Institute of Technology, Bombay in 2012. He received his Ph.D. from the University of Florida in the summer of 2018 During his graduate studies at University of Florida, he worked on several projects with his advisor Dr.Lily Elefteriadou in the Spring of 2016 and 2018