<%BANNER%>

Bare-Earth Extraction and Vehicle Detection in Forested Terrain from Airborne Lidar Point Clouds

Permanent Link: http://ufdc.ufl.edu/UFE0042508/00001

Material Information

Title: Bare-Earth Extraction and Vehicle Detection in Forested Terrain from Airborne Lidar Point Clouds
Physical Description: 1 online resource (154 p.)
Language: english
Creator: Chang, Li-Der
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: bayesian, digital, lidar, morphological, object, support
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Although vehicle detection has been developed for Intelligent Transportation System (ITS), Automatic Vehicle Guidance (AVG), and traffic flow estimation in LiDAR (Light Detection and Ranging) applications, it has not been exploited in cluttered environments such as forested terrain. State-of-the-art airborne LiDAR can provide data in large spatial extents with varying temporal resolution and can be deployed more or less anywhere and anytime, even in cloudy weather and at night. Thus, occluded vehicle detection in forested terrain from airborne LiDAR data can be potentially applied to many fields such as military surveillance, homeland security, global warming, disaster rescue, emergency road service, and criminal searching. In this study, we finished a system with a goal to detect vehicles underneath canopy in forested terrain from LiDAR point clouds. This system covers three important parts and is described as follows. First, the thermal imaging cameras can see the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fog, but not through forest canopy. LiDAR employs an active optical modality or laser ranging that provides primarily geometric information to detect natural surface features and other hard targets that may be spectrally inseparable in multi-spectral passive optical imagery. Accordingly, the first part of this work is that we developed a novel algorithm to help detect obscure targets underneath forest canopy and mitigate the vegetation problem for bare ground point extraction filters as well. By examining the processed results, the forest canopy was successfully removed and all obscure vehicles or buildings underneath canopy can be easily seen. The occluded rate or transparency of forest canopy and the detailed underneath x-y point distribution can be easily obtained accordingly which is very useful for predicting the performance of occluded target detection with respect to various object locations. For the second part of this work, although a variety of algorithms have been developed for extracting the digital terrain model (DTM) from point clouds generated by LiDAR systems, most filters perform well in flat and uncomplicated landscapes, while landscapes containing steep slopes and discontinuities are still problematic. Therefore, we designed a novel bare-Earth extraction algorithm including morphological filtering, segmentation modeling, and surface modeling to automatically classify ground points and non-ground points from LiDAR point clouds. The obtained filtering result has been compared to twelve filters working on the same fifteen study sites which were provided by the ISPRS (International Society for Photogrammetry and Sensing). The average total error and kappa index of agreement of this work in the automated process is 4.6% and 84.5%, respectively, which outperforms all other twelve proposed filters. In addition, we develop another novel slope-based statistical algorithm which is appropriate for any mixed or complicated terrain types. Initially, most objects are removed and initial terrains can be obtained in our object detection algorithm. Slope differences can be assumed to be from a zero-mean normal distribution in all kinds of terrains. Based on slope difference variations, the Chi distribution measurement is used to decide the adaptive slope threshold. Accordingly, the adaptive growing height threshold of each pixel is derived by 8-connected neighbored pixels which can be used to iteratively correct classified points in the initial terrain. The testing results show that this algorithm is even better than our previous algorithm which has outperformed all other twelve algorithms working on the same study sites. In the last part, the obtained canopy-free LiDAR points are clustered into individual objects by the proposed bare-earth extraction algorithm and associated morphological image processing. The clustered LiDAR points of each object are analyzed and classified into the vehicle class or non-vehicle class by many theories such as Spin image, non-parameterric Parzen-window estimation, Bayesian decision, and relative entropy. Finally, the results are demonstrated, discussed, and verified by the Receiver Operating Characteristic (ROC) curve. In addition, we propose another occluded vehicle detection approach, which combines five features extracted from Spin image, PCA (Principal Component Analysis), and LiDAR Intensity (SPI) and applies them to the Support Vector Machine (SVM) classifier. This SPI (Spin image, PCA, and Intensity) method is compared to a simple method and two other vehicle detection methods proposed by other authors papers published by ISPRS. With ten simulations each in different downsampling rates, testing on independent 580 vehicles and 580 non-vehicle objects, our experiments show that this SPI method outperforms all other methods, especially in low sampling rates.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Li-Der Chang.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Harris, John G.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2012-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042508:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042508/00001

Material Information

Title: Bare-Earth Extraction and Vehicle Detection in Forested Terrain from Airborne Lidar Point Clouds
Physical Description: 1 online resource (154 p.)
Language: english
Creator: Chang, Li-Der
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: bayesian, digital, lidar, morphological, object, support
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Although vehicle detection has been developed for Intelligent Transportation System (ITS), Automatic Vehicle Guidance (AVG), and traffic flow estimation in LiDAR (Light Detection and Ranging) applications, it has not been exploited in cluttered environments such as forested terrain. State-of-the-art airborne LiDAR can provide data in large spatial extents with varying temporal resolution and can be deployed more or less anywhere and anytime, even in cloudy weather and at night. Thus, occluded vehicle detection in forested terrain from airborne LiDAR data can be potentially applied to many fields such as military surveillance, homeland security, global warming, disaster rescue, emergency road service, and criminal searching. In this study, we finished a system with a goal to detect vehicles underneath canopy in forested terrain from LiDAR point clouds. This system covers three important parts and is described as follows. First, the thermal imaging cameras can see the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fog, but not through forest canopy. LiDAR employs an active optical modality or laser ranging that provides primarily geometric information to detect natural surface features and other hard targets that may be spectrally inseparable in multi-spectral passive optical imagery. Accordingly, the first part of this work is that we developed a novel algorithm to help detect obscure targets underneath forest canopy and mitigate the vegetation problem for bare ground point extraction filters as well. By examining the processed results, the forest canopy was successfully removed and all obscure vehicles or buildings underneath canopy can be easily seen. The occluded rate or transparency of forest canopy and the detailed underneath x-y point distribution can be easily obtained accordingly which is very useful for predicting the performance of occluded target detection with respect to various object locations. For the second part of this work, although a variety of algorithms have been developed for extracting the digital terrain model (DTM) from point clouds generated by LiDAR systems, most filters perform well in flat and uncomplicated landscapes, while landscapes containing steep slopes and discontinuities are still problematic. Therefore, we designed a novel bare-Earth extraction algorithm including morphological filtering, segmentation modeling, and surface modeling to automatically classify ground points and non-ground points from LiDAR point clouds. The obtained filtering result has been compared to twelve filters working on the same fifteen study sites which were provided by the ISPRS (International Society for Photogrammetry and Sensing). The average total error and kappa index of agreement of this work in the automated process is 4.6% and 84.5%, respectively, which outperforms all other twelve proposed filters. In addition, we develop another novel slope-based statistical algorithm which is appropriate for any mixed or complicated terrain types. Initially, most objects are removed and initial terrains can be obtained in our object detection algorithm. Slope differences can be assumed to be from a zero-mean normal distribution in all kinds of terrains. Based on slope difference variations, the Chi distribution measurement is used to decide the adaptive slope threshold. Accordingly, the adaptive growing height threshold of each pixel is derived by 8-connected neighbored pixels which can be used to iteratively correct classified points in the initial terrain. The testing results show that this algorithm is even better than our previous algorithm which has outperformed all other twelve algorithms working on the same study sites. In the last part, the obtained canopy-free LiDAR points are clustered into individual objects by the proposed bare-earth extraction algorithm and associated morphological image processing. The clustered LiDAR points of each object are analyzed and classified into the vehicle class or non-vehicle class by many theories such as Spin image, non-parameterric Parzen-window estimation, Bayesian decision, and relative entropy. Finally, the results are demonstrated, discussed, and verified by the Receiver Operating Characteristic (ROC) curve. In addition, we propose another occluded vehicle detection approach, which combines five features extracted from Spin image, PCA (Principal Component Analysis), and LiDAR Intensity (SPI) and applies them to the Support Vector Machine (SVM) classifier. This SPI (Spin image, PCA, and Intensity) method is compared to a simple method and two other vehicle detection methods proposed by other authors papers published by ISPRS. With ten simulations each in different downsampling rates, testing on independent 580 vehicles and 580 non-vehicle objects, our experiments show that this SPI method outperforms all other methods, especially in low sampling rates.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Li-Der Chang.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Harris, John G.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2012-12-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0042508:00001


This item has the following downloads:


Full Text

PAGE 1

1 BARE-EARTH EXTRACTION AND VEHICLE DETECTION IN FORESTED TERRAIN FROM AIRBORNE LIDAR POINT CLOUDS By LI-DER CHANG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010

PAGE 2

2 2010 Li-Der Chang

PAGE 3

3 To Families, Teachers and Friends

PAGE 4

4ACKNOWLEDGMENTS First of all, I thank my advisors, Dr. K. Clint Slatton and Dr. John G. Harris, for their great inspiration, support and guidance over my graduate studies. I was impressed by their active thoughts and appreciated very much their supervision which gave me a lot of opportunities to explore my research. I am also grateful to all the members of my advisory committee: Dr. Jose C. Principe, Dr. Dapeng O. Wu, and Dr. Bon A. Dewitt for their valuable time and in terest in serving on my supervisory committee, as well as their comments, which helped improve the qu ality of this dissertation. I also express my appreciation to all the ASPL colleagues, especially Carolyn Krekeler, Karthik Nagarajan, Heezin Lee, Pang-Wei Liu, Kittipat Kampa, Hyun-Chong Cho, an d Sweungwon Cheung for their help, collaboration and valuable discussions during my PhD study. Finally I express my great love for my wife, Jui-Ling Yang and our lovely son, Andrew (Ren-Yu). I thank Jui-Ling for her love, caring, and patience, which made this study possible. Also, I am grateful to my parents for their great support for my life.

PAGE 5

5TABLE OF CONTENTS page ACKNOWLEDGMENTS.............................................................................................................4 LIST OF TABLES................................................................................................................. .......8 LIST OF FIGURES................................................................................................................ .....10 ABSTRACT....................................................................................................................... .........15 CHAPTER 1 INTRODUCTION...............................................................................................................18 Airborne Light Detection and Ranging (LiDAR) Applications..........................................18 Motivation..................................................................................................................... ......19 Organization................................................................................................................... .....20 2 PAPER SURVEYS.............................................................................................................22 Bare-Earth Extraction.......................................................................................................... 22 Morphological Filtering............................................................................................22 Segmentation Modeling............................................................................................22 Surface Modeling......................................................................................................23 Vehicle Detection.............................................................................................................. ..24 Intelligent Transportation System (ITS)....................................................................25 Automatic Vehicle Guidance (AVG)........................................................................25 Traffic Flow Estimation............................................................................................26 3 TEST DATASETS..............................................................................................................27 International Society for Photogrammetry and Sensing (ISPRS) Commission III.............27 City Sites...................................................................................................................27 Forest Sites................................................................................................................27 University of Florida (UF) Airborne Laser Swath Mapping (ALSM) System...................27 City Site.....................................................................................................................2 8 Forest Sites................................................................................................................28 UF Campus Site.........................................................................................................28 4 TREE CANOPY REMOVAL DESIGN.............................................................................33 Multiple Return Analysis....................................................................................................33 Morphological Filtering......................................................................................................35 Tree Canopy Point Decision...............................................................................................36

PAGE 6

6Tree Canopy Removal Result.............................................................................................37 Summary........................................................................................................................ .....38 5 POINT-BASED BARE-EARTH EXTRACTION DESIGN..............................................45 Segmentation Modeling......................................................................................................45 Triangle Classification and Assimilation ................................................................45 Edge Classification and Clustering...........................................................................47 Point Classification....................................................................................................48 Surface Modeling............................................................................................................... .49 Phase-I Point Reclassification Based on Digital Terrain Model (DTM)..................50 Phase-II Point Reclassification Base d on Digital Surface Model (DSM).................51 Phase-III Point Reclassificat ion Based on DTM Roughness....................................52 Phase-IV Point Reclassification Based on Flattened DSM.......................................53 Point-Based Bared-Earth Extraction Result........................................................................54 Same Parameters for All 15 Sites of ISPRS..............................................................54 Optimized Parameters for All 15 Sites of ISPRS......................................................57 Summary........................................................................................................................ .....58 6 GRID-BASED BARE-EARTH EXTRACTION DESIGN................................................75 Object Detection............................................................................................................... ...75 Outlier and Tree Canopy Removal ..........................................................................75 Segmentation by Non-Flat Regions..........................................................................76 Classification of flat regions......................................................................................77 Statistical LiDAR Filtering.................................................................................................78 Chi-Distribution Measurement ................................................................................78 Adaptive Height Threshold Derivation.....................................................................79 Grid-Based Bared-Earth Extraction Result.........................................................................80 Summary........................................................................................................................ .....80 7 OCCLUDED VEHICLE DETECTION DESIGN..............................................................94 Object Clustering.............................................................................................................. ...95 Horizontal Based Morp hological Filtering................................................................96 Vertical Based Morphological Filtering....................................................................97 Object Clustering Result...........................................................................................98 Object Classification.......................................................................................................... .98

PAGE 7

7Principal Component Analysis (PCA)......................................................................99 Spin Image...............................................................................................................100 Bayesian Decision...................................................................................................103 Relative Entropy......................................................................................................104 Occluded Vehicle Detection Result and Evaluation.........................................................106 Summary........................................................................................................................ ...110 8 DOWNSAMPLED VEHICLE DETECTION SIMULATION........................................122 Vehicle and Clutter Dataset...............................................................................................122 Support Vector Machine (SVM)........................................................................................122 Linear SVM.............................................................................................................122 Multi-Class SVM....................................................................................................124 Nonlinear SVM.......................................................................................................124 Novel Feature Extraction..................................................................................................125 Spin Image Features................................................................................................125 Principal Component Features................................................................................126 Surface Intensity Feature.........................................................................................126 Vehicle Detection Methods...............................................................................................128 Vehicle Recognition #1 Method..............................................................................128 Vehicle Recognition #2 Method..............................................................................129 SVM with SPI (Spin image, PC A, and Intensity) Features.....................................130 Downsampled Vehicle Detection Test...............................................................................130 Testing SVM with SPI Features on Hogtown forest sites.................................................131 Summary........................................................................................................................ ...131 9 CONCLUSIONS AND C ONTRIBUTIONS ..................................................................143 Conclusions.................................................................................................................... ...143 Tree Canopy Removal.............................................................................................143 Bare-Earth Extraction..............................................................................................143 Occluded Vehicle Detection....................................................................................145 Contributions.................................................................................................................. ...145 LIST OF REFERENCES..........................................................................................................148 BIOGRAPHICAL SKETCH....................................................................................................154

PAGE 8

8LIST OF TABLES Table page 3-1 Special features of the city sites used in the International Society for Photogrammetry and Sensing (ISPRS) test.....................................................................30 3-2 Special features of the forest sites used in ISPRS test....................................................30 5-1 Extracted and reference cross matrix..............................................................................59 5-2 Interpretation of the kappa index of agreement...............................................................59 5-3 Total errors and kappa index for 15 study sites with same parameters (TH1=1.6m and TH2=0.6m) in our filter........................................................................59 5-4 Total errors and kappa index of agre ement for 15 study sites in our filter with two optimized parameters, TH1 and TH2.......................................................................59 6-1 Associated z positions in the slope formula....................................................................82 6-2 Kernel weights of the smoothing low-pass filter.............................................................82 6-3 Related positions of 8-co nnected neighbored pixels.......................................................82 6-4 Variable list for the 8-connected pixels...........................................................................82 6-5 Total errors and kappa index for 15 study sites by the slop e-based statistical algorithm (SSA)..............................................................................................................82 6-6 Performance comparison among different methods........................................................83 7-1 Summary of some morphologi cal filtering operations..................................................111 7-2 Vehicle detection cross matrix.......................................................................................111 7-3 Vehicle detection accuracy for Hogt own parking site with different Light Detection and Ranging (LiDAR) scans a nd overlapped all LiDAR scans....................111 7-4 Vehicle detection accuracy for Hogtown forest site with different LiDAR scans and overlapped all LiDAR scans...................................................................................111 7-5 Vehicle detection accuracy for Hogtow n residential site with different LiDAR scans and overlapped all LiDAR scans.........................................................................112 8-1 Confusion table of the envelope box classifier.............................................................133 8-2 Confusion table of the SVM (Suppor t Vector Machine) classification by spin-image features........................................................................................................133 8-3 Confusion table of the SVM classifica tion by features (block ing length, blocking width)......................................................................................................................... ....133

PAGE 9

98-4 Confusion table of the SVM classifica tion by features (Surface Intensity Index (SII), blocking area)......................................................................................................133 8-5 Confusion table of the SVM classi fication by features of the vehicle recognition #1................................................................................................................1 33 8-6 Confusion table of the SVM classi fication by features of the vehicle recognition #2................................................................................................................1 33 8-7 Confusion table of the SVM classificatio n by features of SPI (Spin image, PCA, and Intensity) method....................................................................................................133 8-8 Average vehicle detection performan ce of Vehicle Recognition #1 method with 10 simulations for each downsample rate.....................................................................134 8-9 Average vehicle detection performan ce of Vehicle Recognition #2 method with 10 simulations for each downsample rate.....................................................................134 8-10 Average vehicle detection performance of SPI method with 10 simulations for each downsample rate...................................................................................................134 8-11. Vehicle detection accuracy for Hogt own parking site with different LiDAR scans and overlapped all LiDAR scan s based on SVM with SPI features....................135 8-12 Vehicle detection accuracy for Hogtown forest site with different LiDAR scans and overlapped all LiDAR scans based on SVM with SPI features.............................135 8-13 Vehicle detection accuracy for Hogtow n residential site with different LiDAR scans and overlapped all LiDAR scans.........................................................................135

PAGE 10

10LIST OF FIGURES Figure page 3-1 The multiple-return training site on a Google Earth image (screen grab).......................31 3-2 The forest test sites marked on a Google Earth image (screen grab)..............................31 3-3 Images for the University of Florida (UF) campus site...................................................31 3-4 Normalized digital surface model (nDSM) in the UF campus site.................................32 4-1 Tree-canopy removal design flowchart...........................................................................39 4-2 LiDAR (Light Detection and Ranging) multiple return shot illustration........................39 4-3 Snapshot image of the site22 from the Google Earth......................................................40 4-4 Histogram of shot-wise height differences at the site22.................................................40 4-5 Zoom in the smaller difference part of Figure 4-4..........................................................41 4-6 Shot-wise height difference map c lipped to 0.0m-0.1m at the site22.............................41 4-7 Multiple return map at the samp22 city site....................................................................42 4-8 Morphological filtering result s at the samp22 city site...................................................42 4-9 Detected vegetation area (canopy cell) ma rked on DSM of the city test site.................43 4-10 Canopy removal result for the city test site.....................................................................43 4-11 Revealed occluded vehicles at the Hogtown parking site...............................................43 4-12 Revealed occluded vehicles at the Hogtown forest site..................................................44 4-13 Revealed occluded vehicles at the Hogtown residential site...........................................44 4-14 Remained point density in 2-D di stribution after removing canopy points....................44 5-1 Segmentation modeling flowchart of our bare-earth extraction design..........................60 5-2 Triangle classification a nd assimilation illustration........................................................60 5-3 Triangle assimilation curve of samp22 site.....................................................................60 5-4 Cluster-wise ground ratio illustration..............................................................................61 5-5 Vector-based edge classification.....................................................................................61 5-6 Cluster-wise boundary edge detection............................................................................61 5-7 Initial point classification result......................................................................................62 5-8 Surface modeling flowchart of our bare-earth extraction design....................................62 5-9 Phase-I point reclassification result.................................................................................63 5-10 Roughness map of canopy removal DSM.......................................................................63 5-11 Phase-II point re classification result...............................................................................63 5-12 Roughness map of DTM (Digital Terrain Model)...........................................................64

PAGE 11

115-13 Bridge detection result................................................................................................... ..64 5-14 Point reclassification phase-III result..............................................................................64 5-15 The flattened DSM map..................................................................................................65 5-16 Final point cla ssification result........................................................................................6 5 5-17 DSM image of the city site #1.........................................................................................65 5-18 The final classification errors in the site of samp11........................................................66 5-19 The final classification errors in the site of samp12........................................................66 5-20 DSM image of the city site #2.........................................................................................66 5-21 The final classification errors in the site of samp21........................................................67 5-22 The final classification errors in the site of samp22........................................................67 5-23 The final classification errors in the site of samp23........................................................67 5-24 The final classification errors in the site of samp24........................................................68 5-25 DSM image of the city site #3.........................................................................................68 5-26 The final classification errors in the site of samp31........................................................68 5-27 DSM image of the city site #4.........................................................................................69 5-28 The final classification errors in the site of samp41........................................................69 5-29 The final classification errors in the site of samp42........................................................69 5-30 DSM image of the forest site #5......................................................................................70 5-31 The final classification errors in the site of samp51........................................................70 5-32 The final classification errors in the site of samp52........................................................70 5-33 The final classification errors in the site of samp53........................................................71 5-34 The final classification errors in the site of samp54........................................................71 5-35 DSM image of the forest site #6......................................................................................71 5-36 The final classification erro rs in the site of samp61........................................................72 5-37 DSM image of the forest site #7......................................................................................72 5-38 The final classification erro rs in the site of samp71........................................................72 5-39 Accuracy comparison to the best two filters for all 15 study sites..................................73 5-40 Accuracy comparison of our automa ted filter to optimized filter...................................73 5-41 Average accuracy comparison to ei ght different filters before 2004..............................74 5-42 Average accuracy comparison to four different filters proposed after 2004...................74 6-1 Standard deviation and confidence interval of a normal distribution.............................84 6-2 Outlier points found by a normal distribution at site22...................................................84

PAGE 12

126-3 Edge detection at the site22.............................................................................................84 6-4 Obtained flat regi ons at the site22...................................................................................85 6-5 Rule-based methods for fl at region classification...........................................................85 6-6 The associated feature maps and classification result of the site22................................86 6-7 DTM comparison at the site22........................................................................................87 6-8 DTM comparison at the site11........................................................................................87 6-9 DTM comparison at the site12........................................................................................87 6-10 DTM comparison at the site21........................................................................................88 6-11 DTM comparison at the site23........................................................................................88 6-12 DTM comparison at the site24........................................................................................88 6-13 DTM comparison at the site31........................................................................................89 6-14 DTM comparison at the site41........................................................................................89 6-15 DTM comparison at the site42........................................................................................89 6-16 DTM comparison at the site51........................................................................................90 6-17 DTM comparison at the site52........................................................................................90 6-18 DTM comparison at the site53........................................................................................90 6-19 DTM comparison at the site54........................................................................................91 6-20 DTM comparison at the site61........................................................................................91 6-21 DTM comparison at the site71........................................................................................91 6-22 Histograms of slope differences from the initial terrain at the site22.............................92 6-23 Adaptive slope difference threshold algorithm based on Chi distribution......................93 6-24 The change curve of freedom at the site22......................................................................93 6-25 Histogram of remaining slope differences at the site22..................................................93 7-1 Object segmentation flowchart......................................................................................113 7-2 Object segmentation result in Hogtown parking site....................................................113 7-3 Object segmentation result in Hogtown forest site.......................................................113 7-4 Object segmentation result in the residential site..........................................................114 7-5 Object classification flowchart......................................................................................114 7-6 Using Principal Component Analysis (P CA) to get the best fitting plane....................115 7-7 An unoccluded vehicle consis ting of 311 LiDAR points..............................................115 7-8 An occluded vehicle consisting of 82 LiDAR points....................................................115 7-9 An occluded vehicle consisting of 73 LiDAR points....................................................116

PAGE 13

137-10 An occluded vehicle consisting of 69 LiDAR points....................................................116 7-11 An occluded vehicle consisting of 67 LiDAR points....................................................116 7-12 An occluded vehicle consisting of 41 LiDAR points....................................................116 7-13 An occluded vehicle consisting of 14 LiDAR points....................................................117 7-14 An occluded vehicle consisting of 12 LiDAR points....................................................117 7-15 An occluded vehicle consisting of 8 LiDAR points......................................................117 7-16 An occluded vehicle consisting of 7 LiDAR points......................................................117 7-17 Bivariate Probability Density Func tion (PDF) of reference vehicles............................118 7-18 Bivariate PDF of reference non-vehicles......................................................................118 7-19 The information divergence for vehicles and non-vehicles..........................................118 7-20 Bivariate PDF of detected vehicles with BT=0.7..........................................................119 7-21 Bivariate PDF of detected non-vehicles with BT=0.7..................................................119 7-22 Occluded vehicle detection results in the Hogtown parking site..................................119 7-23 Occluded vehicle detection result s in the Hogtown forest site.....................................120 7-24 Occluded vehicle detection results in the Hogtown residential site..............................120 7-25 Vehicle detection performance vs. th e number of collected LiDAR points..................120 7-26 Vehicle detection accuracy vs. the number of collected LiDAR points........................121 7-27 Receiver Operator Characteristics (ROC) curves of our detection performance..........121 8-1 LiDAR point density histogram of refe rence vehicles in the UF campus site..............136 8-2 A DSM and LiDAR point map example.......................................................................136 8-3 Spin image of Figure 8-2...............................................................................................136 8-4 SVM (Support Vector Machine) classifi cation result by Spin-image features..............137 8-5 Length and width estimati on by PCA in Figure 8-2......................................................137 8-6 SVM classification result by blocki ng length, blocking wi dth features........................138 8-7 Intensity Map gridded from LiDAR data at one of the UF campus site.......................138 8-8 Point relationship with the convex hull.........................................................................138 8-9 SVM classification result by surface intens ity index and blocking area features.........139 8-10 The six-parameter representation of the vehicle recognition #1 method......................139 8-11 Contained information of PCA of the vehicle recognition #1 method..........................140 8-12 Contained information of PCA of the vehicle recognition #2 method..........................140 8-13 Contained information of PCA of the novel features....................................................140 8-14 Average recognition rate of vehi cle detection vs. downsample rate.............................141

PAGE 14

148-15 Average Kappa index of agreement of vehicle detection vs. downsample rate............141 8-16 Contained information of principal co mponents of the SPI (Spin image, PCA, and Intensity) method applied to Hogtown forest sites.................................................142

PAGE 15

15Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy BARE-EARTH EXTRACTION AND VEHICLE DETECTION IN FORESTED TERRAIN FROM AIRBORNE LIDAR POINT CLOUDS By Li-Der Chang December 2010 Chair: John G. Harris Major: Electrical and Computer Engineering Although vehicle detection has been develope d for Intelligent Transportation System (ITS), Automatic Vehicle Guidance (AVG), and traffic flow estimation in LiDAR (Light Detection and Ranging) applications, it has not been exploited in cluttered e nvironments such as forested terrain. State-of-the-art airborne LiDAR can provide data in large spatial ex tents with varying temporal resolution and can be deployed more or less anywhere and anytime, ev en in cloudy weather and at night. Thus, occluded vehicle detection in forested terrain from airborne Li DAR data can be potentially applied to many fields such as military surveillance, homeland security, gl obal warming, disaster rescue, emergency road service, and criminal searching. In this study, we finished a system with a goal to detect vehicles underneath canopy in forested terrain from LiDAR poi nt clouds. This system covers three important parts and is described as follows. First, the thermal imaging cameras can see the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fo g, but not through forest canopy. LiDAR employs an active optical modality or laser ranging that provid es primarily geometric information to detect natural surface features and other hard targets that may be sp ectrally inseparable in multi-spectral passive optical imagery. Accordingly, the first part of this work is th at we developed a novel algorithm to help detect obscure targets underneath forest canopy and mitigat e the vegetation problem for bare ground point extraction filters as well. By examining the pro cessed results, the forest canopy was successfully

PAGE 16

16 removed and all obscure vehicles or buildings u nderneath canopy can be easily seen. The occluded rate or transparency of forest canopy and the detailed underneath x-y point distribution can be easily obtained accordingly which is very useful for predicting the pe rformance of occluded target detection with respect to various object locations. For the second part of this work, although a va riety of algorithms have been developed for extracting the digital terrain model (DTM) from point clouds generated by LiDAR systems, most filters perform well in flat and uncomplicated landscap es, while landscapes containing steep slopes and discontinuities are still problematic. Therefore, we designed a novel bare-Earth extraction algorithm including morphological filtering, segmentation modeling, and surface modeling to automatically classify ground points and non-ground points from LiDAR point clouds. The obtained filtering result has been compared to twelve filters working on the sam e fifteen study sites which were provided by the ISPRS (International Society for Photogrammetry and Sensing). The average total error and kappa index of agreement of this work in the automa ted process is 4.6% and 84.5%, respectively, which outperforms all other twelve proposed filters. In addition, we develop another novel slope-based statistical algorithm which is appropriate for any mi xed or complicated terrain types. Initially, most objects are removed and initial terrains can be obtai ned in our object detection algorithm. Slope differences can be assumed to be from a zero-mean normal distribution in all kinds of terrains. Based on slope difference variations, the Chi distribution me asurement is used to decide the adaptive slope threshold. Accordingly, the adaptive growing height threshold of each pixel is derived by 8-connected neighbored pixels which can be used to iteratively correct classified points in the initial terrain. The testing results show that this algorithm is ev en better than our previous algorithm which has outperformed all other twelve algorithms working on the same study sites. In the last part, the obtained canopy-free LiDAR poi nts are clustered into individual objects by the proposed bare-earth extraction algorithm and associ ated morphological image processing. The clustered LiDAR points of each object are analyzed and classified into the vehicle class or non-vehicle class by many theories such as Spin image, non-parametric Parzen-window estimation, Bayesian decision, and

PAGE 17

17 relative entropy. Finally, the results are demonstrat ed, discussed, and verified by the Receiver Operating Characteristic (ROC) curve. In addition, we propose another occluded vehicle detection approach, which combines five features extracted from Spin image, PCA (Principal Component Analysis), and LiDAR Intensity (SPI) and applies them to the Suppor t Vector Machine (SVM) classifier. This SPI (Spin image, PCA, and Intensity) method is compared to a simple method and two other vehicle detection methods proposed by other authors’ papers published by ISPRS. With ten simulations each in different downsampling rates, testing on independent 580 ve hicles and 580 non-vehicle objects, our experiments show that this SPI method outperforms all other methods, especially in low sampling rates.

PAGE 18

18CHAPTER 1 INTRODUCTION Airborne light detection and ranging (LiDAR) te chnology is an active remote sensing technology which allows accurate measurements of topography, vegetation canopy heights, and buildings over large areas. Most modern ALTM systems consist of thr ee basic components: the laser scanner, a kinematic Global Position System (GPS), and the Inertial Measurement Unit (IMU). The laser scanner detects the range from aircraft to ground by recording the time difference between laser pulses sent out and reflected back. In this study, we used LiDAR point clouds to develop an automatic canopy removal algorithm and a novel bare-Earth extraction algorithm to reveal those LiDAR points underneath forest canopy and filter ground points, respectively. Accordingly, the goal of o ccluded vehicle detection in forested terrain can be continually exploited and achieved. Airborne LiDAR Applications In the last decade, advances in airborne LiDAR technologies have enabled dramatic improvements in topographic mapping resolution, particularly for ground surfaces beneath vegetation canopies [1], [2]. LiDAR data analyses have proven successful for a variet y of forested terrain applications, such as fusing multi-resolution elevation measurements [3], estimation of bare-surface topography [4], [5], extraction of micro-stream channels beneath forest canopies [6], es timation of sunlight flux in forest understories [7], segmentation of individual tree canopies [8], and stand-level forest parameter estimation [9]. A DTM (Digital Terrain Model), commonly used in terchangeably with a digital elevation model (DEM), is a digital representation consisting of terra in elevations for ground positions. A DTM is also called a bare-Earth model since it excludes features on the Earth such as tall vegetation, buildings, and bridges. A DTM is generally acquired using fi eld-based land surveying methods, photogrammetric techniques, or through processing remotely sensed data. Airborne LiDAR, an active remote sensing technology, has revolutionized the acquisition of digital elevation data for large-scale mapping applications and has become a fixture of presentday mapping missions, provi ding cost-effective means to achieve high-accuracy results.

PAGE 19

19 Accordingly, Airborne LiDAR can produce detaile d maps of the ground more speedily and, in many cases, more economically than almost any other method. Using LiDAR will result in a DTM of the ground as well as the top (height) of the vegetation so that many features can also be determined, especially in conjunction with standard or sma ll frame aerial photography (such as access roads, structures and water courses). The result is that these data can be used to help assess stands and timber volume, plan roads, review slope stability, run-o ff and other critical data about any area which are subject to flood and drainage modeling, land-use studies, geological applications, as well as urban planning and management. Motivation ISPRS (International Society for Photogrammetry and Sensing) Working Group III/3 conducted a test [10] and found th at all bare ground point extractio n filter algorithms perform well on LiDAR point clouds from smooth rural landscapes, but all produce errors in rough terrain with vegetation canopy. Besides, the thermal imaging camer as can see the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fog, but not through forest canopy. LiDAR employs an active optical modality or laser ranging that provides primarily geometric information to detect natural surface features and other hard targets that may be spectrally inseparable in multi-spectral passive optical imagery. Therefore, th e first part of this work is to develop a novel algorithm to mitigate the vegetation problem for tho se ground point extraction filters and help detect obscure targets underneath forest canopy as well. In the last decade, a variety of algorithms have been developed for extracting DTMs from point clouds generated by LiDAR systems, where automa tic and robust ground point extraction has been attracting a great deal of attention. Three kinds of approaches for LiDAR filtering are most prevalent: morphological filtering, segmentation modeling, and surface modeling. Although most filters perform well in flat and uncomplicated landscapes, landscapes containing steep slopes and discontinuities are still problematic. Hence, the second part of this work is to design a novel bare-earth extraction algorithm

PAGE 20

20 which can combine those three prevalent approach es and to achieve better performance than other existing algorithms. In addition, airborne LiDAR can provide data in large spatial extents with varying temporal resolution and it can be deployed more or less anywhe re and at anytime, including in smoke, haze, fog, and nighttime, which restrict the use of passive op tical imagery. Thus, occluded vehicle detection from airborne LiDAR data in forested terrain, the third part of this work, can be applied to many fields, more specifically: 1) military surveillance – searching for enemy vehicles in a battle area with forest, 2) homeland security – border crossing monitoring for vehi cles in forest area, 3) global warming – vehicle hunting for illegal deforestation which is a hidden cause of global warming, 4) disaster rescue – finding vehicles stuck by disrupted roads in forest during natural disaster, 5) emergency road service – locating vehicles involved with general car accidents in forest, and 6) criminal searching – uncovering forest canopy to search suspicious vehicles hiding in mountains. Organization This dissertation is organized as follows. First we re view some papers associated with this study which include bare-Earth extraction and vehicle detection. Morphological filtering, segmentation modeling, and surface modeling, which are three more prevalent approaches for LiDAR filtering algorithms, are briefly discussed. Intelligent transpor tation system (ITS), automatic vehicle guidance (AVG), and traffic flow estimation, which are rela ted to vehicle detection research, are introduced and summarized. In Chapter 3, we explain test datasets of LiDAR point clouds in this study which were provided by ISPRS Commission III and UF (Universit y of Florida) Airborne Laser Swath Mapping (ALSM) system. Both city sites and forest sites are investigated. In Chapter 4, we show our tree canopy removal design which consists of multiple return an alysis, morphological filtering, and tree canopy point decision. The results are also demonstrated and discu ssed. In Chapter 5, we show our novel bare-earth extraction design which is composed of two sections. The first section, segmentation modeling, is formed by triangle classification and assimilation, edge classifi cation and clustering, and point classification. The second section, surface modeling, is constructed by phase-I point reclassification based on DTM,

PAGE 21

21 phase-II point reclassification based on DSM (D igital Surface Model) roughness, phase-III point reclassification based on DTM roughness, and phase-IV point reclassification based on flattened DSM. The results are also demonstrated and discussed. In addition, a slope-based statistical bare-Earth extraction algorithm is proposed in Chapter 6. Obj ect detection is achieved by slope relationships between clustered points and morphological filtering. With detected and removed object points, initial ground points are obtained and reclassified by compar ing the above ground level to the adaptive height threshold derived from Chi distribution measuremen t. The occluded vehicle detection designs are exploited in Chapter 7 and Chapter 8. In Chapter 7, the object clustering is completed by horizontal based and vertical based morphological filtering, where the results are also presented. Object classification is followed and achieved by Spin im age, non-parametric Parzen-window estimation, Bayesian decision, and relative entropy. In Chapter 8, we propose to extract features from Spin image, principal component analysis, and LiDAR intensity an d apply them to the support vector machine for detecting vehicles. The testing database includ es 580 independent vehicles in open area and 580 independent non-vehicles in clutte red and occluded environment. Th e occluded vehicle detection is simulated by detecting those objects from randomly downsampling LiDAR points of vehicles mixing with cluttered objects. Finally, Chapter 9 is our conclusions and contributions.

PAGE 22

22CHAPTER 2 PAPER SURVEYS Bare-Earth Extraction There are many existing approaches for filtering Li DAR to classify LiDAR points into ground and object points for DTM generation and further reconstr uction of topographic features. Three kinds of approaches for LiDAR filtering are more prevalent: morphological filtering, segmentation modeling, and surface modeling. Although most filters perform we ll in flat and uncompli cated landscapes, landscapes containing steep slopes and discontinuities are still problematic. Morphological Filtering In this first kind of approach for LiDAR filte ring, the structure element of mathematical morphological filtering was used by Vosselman [11], in which the admissible height difference for ground points depends on the horizontal distance be tween a ground point and its neighboring points. A larger neighboring horizontal distance will allo w a greater height difference among accepted ground points. This structure element is placed at each poi nt, so ground points are identified as those that fall below the admissible height difference. Using multiple structure elements was also proposed by Kilian [12], where the likelihood of points as ground is weighted by each window size and then ground points are determined accordingly in the final classification. Zhang [13] proposed a progressive morphological filter to detect non-ground features. By gradually increasing the window size of the filter and using elevation difference thresholds, the measurements of vehicles, vegetation, and buildings are removed, while ground data are preserved. A similar concept is a slope-based method proposed by Sithole [14], Roggero [15], and Kampa [16]. The slope or height difference between two points is measured. If that measurement exceeds a certain threshold, the higher point is assumed to belong to an object point. Segmentation Modeling In this second kind of approach for LiDAR filtering, points are aggregated into segments by the geometric relationship of neighborhoods based on hei ght, slope or curvature difference. Sithole [17] proposed a method for object segmentation, where a point cloud is sliced into parallel vertical profiles

PAGE 23

23 and aggregated based on proximity to determine the adjacency of surface segments. The purpose of segmentation is to obtain higher level information from points in a point cloud, which is usually knowledge of the extent of homogeneous regions in a landscape, e.g., buildings, vegetation, and bridges. Nardinocchi [18] presented a strategy for the classifi cation of raw LIDAR data as terrain, buildings, and vegetation. Guided by that classification, the si ze and relationship of adjacent segmented objects are analyzed in order to get final point classification. F ilin [19] proposed a clustering analysis in which the position, the best fitting plane parameters, and he ight difference of neighboring points are used. Vosselman [20] also proposed the use of the Hough tran sformation to detect planar roof surfaces within the given building boundaries for clustering. In Brovelli [21], it was proposed that any points belonging to an object must be segmented as a cluster if th ey are above its neighborhood height differences. The primary objectives for segmentation [17] are to obtain better discrimination of large objects in a landscape, preserve discontinuities, and allow both Type I and II errors to be reduced instead of making a trade off between them. Surface Modeling Finally, the filters in the third group are based on a surface model, where the entire point set is updated iteratively to approach the ground surface. The residuals of all points are measured by the relation between their height levels and a surface mo del. Measured points that lie above a surface should have less influence on the shape of the surface, while points lying below should have more influence. In Kraus’s paper [22], this surface runs in an averag ing way between terrain poi nts and vegetation points. The terrain points are more likely to have negativ e residuals, whereas the vegetation points are more likely to have small negative or positive residuals. Elmqvist [23] used an active shape model, also referred to as snakes, for detection of contours in im ages, where the inner forces of the surface determine its stiffness and the external forces are a negative grav ity. Iteration starts with a horizontal surface below all points that moves upward to reach the point, but inner stiffness prevents it from reaching up to the points on vegetation or house roofs. Axelsson [24] used a sparse TIN derived from neighborhood minima as the first reference surface. In each iteration, if a point is found with off sets below given threshold

PAGE 24

24 values, it is classified as a ground point and added to th e TIN such that the TIN is progressively densified. The iterative process stops when no more points are below the threshold. However, an experimental study of some filtering algorithms was described by Sithole in 2004 [25], where their performances were co mpared based on the results of filtering the same LiDAR data sets provided by ISPRS. The comparison results determined that in flat and uncomplicated landscapes (i.e., small to medium sized buildings standing well off a fairly flat ground) algorithms tend to do well. Significant differences in accuracies of filtering appear in landscapes containing steep slopes and discontinuities. These differences are a result of the ability of algorithms to preserve terrain discontinuities while detecting large objects. After the year 2004, Sithole [26], Silvan-Cardenas [27], Lu [28], and Meng [ 29] also used the same data sets to develop new filtering algorithms. However, these filters perform insignificantly worse than the filter developed by Axelsson in 2000 [24]. In th is study, we present a novel filter composed of tree canopy removal design, bare-earth extraction design, and point reclassification design, which includes features of morphological filtering, segmentation modeling, and surface modeling, respectively, described in the above three filtering groups. The fi nal filtering results show that our automated process works better than Axelsson’s filter, and the performance is even better after two dominant parameters are optimized manually. Vehicle Detection Although vehicle detection has been utilized in Intelligent Transportation System (ITS), Automatic Vehicle Guidance (AVG), and traffic flow estimation, it has not been exploited in forested terrain. Airborne LiDAR can provide data in large spatial ex tents with varying temporal resolution and it can be deployed more or less anywhere and anytime, in cluding in smoke, haze, fog, and nighttime, which restrict the use of passive optical imagery. Thus, o ccluded vehicle detection from airborne LiDAR data in forested terrain can be applied to many fields, mo re specifically: 1) military surveillance – searching for enemy vehicles in a battle area with forest, 2) homeland security – border crossing monitoring for vehicles in forest area, 3) global warming – vehicl e hunting for illegal deforestation which is a hidden

PAGE 25

25 cause of global warming, 4) disaster rescue – finding vehicles stuck by disrupted roads in forest during natural disaster, 5) emergency road service – locati ng vehicles involved with general car accidents in forest, and 6) criminal searching – uncovering for est canopy to search suspicious vehicles hiding in mountains. Intelligent Transportation System (ITS) Vehicle detection has been applied to Intelligen t Transportation System and its summary can be reviewed in [30]. The associated technologies for ve hicle detection can be categorized into intrusive sensors and non-intrusive sensors which depend on wh ether the sensors are required to be installed directly onto or into the road surface or not. The intr usive sensors include the pneumatic road tube [31], inductive loop [32], piezoelectric cable [33], and magne tic sensors [34], [35], [36] and weigh-in-motion (WIM) systems [37]. The non-intrusive sensors are m ounted overhead or on the side of the roadway, such as the video image processor [3 8], [39], [40], microwave radar [41], active and passive infrared sensors [40], ultrasonic sensors [42], and passive acous tic array sensors [30]. However, both the intrusive sensors and non-intrusive sensors need installation in appropriate spots which also restrains the ability to detect vehicles everywhere. Automatic Vehicle Guidance (AVG) Vehicle detection with vision-based methods h as been devoted to Automatic Vehicle Guidance (AVG) [43]. A review of on-road vehicle detection ca n be found in [44]. So far, several prototypes and solutions have been produced based on different appr oaches [45], [46], [47], [48]. Looking at research on intelligent vehicles worldwide, Europe pioneers the research, followed by Japan and the United States. Two kinds of sensors, passive type and active type are mounted on the experimental vehicle and exploited to the approach of vehicle detection. For the passive sensors [49], such as normal cameras, their drawbacks are that they are too susceptible to shadows, occlusion, dayto-night transition, and inclement weather which restrict these kinds of sensors to be used at any time. For the active sensors [49], such as radar-based [50], laser-bas ed (i.e. LiDAR) [51], [52], and acoustic-based [53], their performances in fog, rain, or snow are better than passive senso rs. But, a big problem is posed by the interference

PAGE 26

26 among the same types of sensors when a large number of vehicles is moving simultaneously in the same direction. However, that is not a problem for using airborne LiDAR to detect vehicles since the mission for searching vehicles in the region of interest can be completed by the divide and conquer strategy [54]. Traffic Flow Estimation Recently, vehicle detection with airborne LiDAR data is exploited to support traffic flow estimation [55], [56], [57], [58], [59]. In their system arch itecture, the main processing steps are road surface extraction, vehicle extraction, primary parameterizati on of vehicle shape, feature space selection, vehicle classification, vehicle velocity estimates, and traffic flow data computation. However, the road area has to be filtered in the beginning for their following vehicle extraction process. Although the road surface modeling can be derived from air borne LiDAR data [56], they still need the GIS or CAD database to assist in extracting the road edges and road medians. Th erefore, if a vehicle is being driven or parked on some road or place which information is out of da te or unavailable in the associated database, it will cause missing errors for vehicle detection. In additio n, the detection for those occluded vehicles which are underneath the trees or forest is out of their scope.

PAGE 27

27CHAPTER 3 TEST DATASETS ISPRS Commission III City Sites ISPRS Commission III/WG3 provides LiDAR data sets including both urban landscapes (city sites) and rural landscapes (forest sites) which were acqui red by an Optech Airborne Laser Terrain Mapper (ALTM) scanner over the Vaihingen/Enz test field and the Stuttgart city center [57]. The average point density for the city sites is roughly 0.67 points per square meter. Every record in the raw data represents three-dimensional coordinates and in tensity of first and last returned points, while each record in the reference data has been compiled by semi-automa tic filtering and manual editing as either a ground or non-ground point. These areas were chosen because of th eir diverse feature content such as vegetation, buildings, roads, railroads, ramps, bridges, power lines, etc. Some special features of these sites are listed in Table 3-1. Forest Sites The average point density for the forest sites is r oughly 0.18 points per square meter. Every record in the raw data also represents three-dimensional c oordinates and intensity of first and last returned points. Each record in the reference data has also been compiled by semi-automatic filtering and manual editing as either a ground or non-ground point. Thes e areas include diverse feature content such as vegetation, rivers, ridges, buildings, etc. Some special features of these sites are listed in Table 3-2. The forest site 8 was not used in this resear ch since no reference data was available. UF ALSM System The UF airborne LiDAR system, purchased in 2007, is a state-of-the-art commercial LiDAR (Optech Gemini) offering laser pulse rates of up to 16 7 kHz, with up to 4 returns per transmitted pulse, recorded intensities of each return, and multiple beam divergences [58]. For this study, the LiDAR data sets were collected in Dec 2008 by the Optech Gemini operated at 125 KHz pulse rate, 45 Hz scan frequency, 0.3 mrad beam divergence, 213 Km/h airplane speed, 0 to 15 scan angle, and 600m Above

PAGE 28

28 Ground Level (AGL) for 14 repeat passes around the Hogtown area and UF campus with vertical accuracy of about 15 cm and the average LiDAR density is 3.52 shots per square meter. City Site In order to further investigate and verify the mu ltiple-return characteristic of airborne LiDAR data, the site close to the UF campus was introduced and collected by the UF airborne LiDAR system. This site’s image snapshot from Google Earth is shown in Figure 3-1.This area was chosen because of its diverse feature content including uncovered ground, still vehicles, light poles, buildings, and isolated trees. Besides, the average LiDAR point density is 4.44 points per square me ter which can give better rendering resolution than the data set provided by ISPR S. Since the edges of small or large objects and discontinuity of trees have less probability to be scanned and reflected, the higher LiDAR density can provide more detailed information and ma ke the multiple-return analysis easier. Forest Sites Three distinct forest sites in Hogtown area were collected by the UF airborne LiDAR system and used in this research. The Hogtown area is a mixed warm temperate forest composed of roughly 80 percent deciduous and 20 percent coniferous trees, but this mix varies significantly with location. Hogtown consists of trees as tall as 35 meters and possesses the characteristics expected in a natural forest where there are multiple layers of folia ge, varying and random spacing between trees, and significant undergrowth. The locations of 3 survey si tes (Hogtown forest site, Hogtown parking site, and a residential site) close to the UF campus are ma rked on Google Earth and labeled in Figure 3-2. UF Campus Site The LiDAR scanning data of one flight around the no rth-west corner of the UF campus is selected for downsampled vehicle detection simulation. Th e photo image, LiDAR digital surface model (DSM) and DTM of this area are showed in Figure 3-3. We are interested in this area because it covers many unoccluded vehicles and trees. In order to see those object s clearer in this area, the partial data in a parking lot and a residential site are cropped which ar e marked by squares in the Figure 3-3b and Figure

PAGE 29

29 3-3c. The normalized digital surface models (nDSM) are shown in Figure 3-4 which are calculated by using their DSM minus DTM to get relative heights of those objects.

PAGE 30

30 Table 3-1. Special features of the city sites used in ISPRS test Site Reference data Special features City site1 Samp11 Buildings on steep slopes City site1 Samp12 Vegetati on mixed with buildings City site2 Samp21 Bridge and large building City site2 Samp22 Bridges & terrain discontinuities City site2 Samp23 Irregularly shaped buildings City site2 Samp24 Ramp & terrain discontinuities City site3 Samp31 Large buildings City site4 Samp41 Data gaps and outliers City site4 Samp42 Railway station with trains Table 3-2. Special features of the forest sites used in ISPRS test Site Reference Data Special Features Forest site 5 Samp51 Vegetation on steep slopes Forest site 5 Samp52 Vegetation on river bank Forest site 5 Samp53 Terrain discontinuities Forest site 5 Samp54 Low resolution buildings Forest site 6 Samp61 Sharp ridge and embankments Forest site 7 Samp71 Ramp, bridge, and underpass

PAGE 31

31 Figure 3-1. The multiple-return training s ite on a Google Earth image (screen grab) Figure 3-2. The forest test sites marked on a Google Earth image (screen grab) A x (meter)y (meter) 200 400 600 800 1000 1200 1400 1600 1800 2000 50 100 150 200 0 10 20 B x (meter)y (meter) 200 400 600 800 1000 1200 1400 1600 1800 2000 50 100 150 200 0 10 20 C Figure 3-3. Images for the UF campus site. A) Snapshot image from Google Earth, B) LiDAR DSM, and C) LiDAR DTM.

PAGE 32

32 x (meter)y (meter) 25 50 75 100 125 150 175 200 225 25 50 75 100 125 150 175 200 225 0 0.5 1 1.5 2 2.5 3 A x (meter)y (meter) 25 50 75 100 125 150 175 200 225 25 50 75 100 125 150 175 200 225 0 5 10 15 B Figure 3-4. Normalized digital surface model in the UF campus site: A) nDSM of the parking lot and B) nDSM of the urban forest.

PAGE 33

33CHAPTER 4 TREE CANOPY REMOVAL DESIGN This chapter is the first design part of our work First, the multiple-return characteristic of LiDAR data is analyzed and accordingly laser shots are clas sified as single-return or multiple-return shots. The major challenge of removing canopy is that some fo liage will unexpectedly reflect single-return shots rather than normal multiple-return s hots when they are very dense. This challenge can be solved by using our developed algorithms such as analyzing dist ance relationships between foliage, applying morphological filters to process the canopy/non-canopy image and creati ng rough digital terrain models to calculate above ground levels of points, etc. The unique feature of this approach is that no para meter tweaking is required. Both of the city and forest sites are tested where the data are from ISPRS a nd UF, respectively. It shows that all tree or forest canopy points have been removed such that all ob scure vehicles or buildings underneath canopy can now be easily seen. For military application, although ther mal imaging cameras can s ee the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fog, they cannot be used to see through the forest canopy, unlike our algorithm. The canopy removal algorithm, a block diagram of which is shown in Figure 4-1, takes th e multiple-return feature to classify raw point clouds into canopy and non-canopy points throu gh morphological filtering techniques. Multiple Return Analysis At the beginning, our LiDAR datasets consist of one two, three or four return points for each laser shots recorded in terms of x, y, and z coordinat es and reflected intensity. Based on the observation and analysis in the LiDAR scanned data of a shopping plaza near the UF campus, all multiple-return shots occur on tree crowns, bushes, edges of buildings, et c, where multiple-return shots posses two, three or four return points in their individual shots and ar e represented by black dots in the Figure 4-2, in which the area size is about 180m by 130m. It is noted that there are no multiple-return shots occuring on any vehicles. These multiple-return shots are caused due to th e fact that the beam area of a single laser shot covers more than one reflected spot which often happens on small leaves and sharp edges.

PAGE 34

34 After analyzing the observation in Figure 4-2, tw o challenges which cause missing errors and false alarm errors are required to be conquered for removing tree canopy. Th e first challenge is that dense foliage will unexpectedly reflect single-return shots instead of normal multiple-return shots which can be observed on some tree tops in Figure 4-2 and lead to missing errors of tree canopy detection. The second one is that some sharp edges or protrusions of bu ildings and non-tree objects also give multiple-return shots similar to reflection from leaves of trees wh ich will cause false alarm errors. These two problems can be solved by four steps: 1) analyze distan ce property between normal and dense foliage; 2) apply morphological filters including closing, opening, and dilating operators to process the binary canopy/non-canopy image; 3) create rough DTM to calc ulate above ground levels of points; 4) remove all canopy points above the maximum height of vehicles. We also analyzed another city site “samp22” from ISPRS data sets, where its average point density is about 0.91 points per square meter and its image w as snapshot from Google Earth [59] and is shown in Figure 4-3. This area was chosen because of its dive rse feature content such as vegetation, buildings, roads, vehicles, bridges, etc. Its dataset only consists of the first-return and last-return points of laser shots. Although every shot of ISPRS was also recorded in terms of x, y, and z coordinates and reflected intensity, each shot always goes with two returns, fi rst-return and last-return. Under this situation, one more procedure is required to categorize each shot in to either the multiple-return class or single-return class. By calculating the difference between the z coor dinates of the first-return point and last-return point of each shot, all shot-wise vertical distances ar e immediately obtained. By analyzing the histogram of shot-wise height differences in Figure 4-4 at the site samp22, it is found that multiple-return shots and single-return shots are in two separate groups which ar e either above or below 1 meter. If we zoom in on the smaller difference part and set all differences la rger than 1 meter equal to 1 meter, shown as Figure 4-5, most single-return shots can be classified by a simple threshold of 10 cm. This threshold covers 99% of single-return shots, while the rest of them can be considered outlier situations. Additionally, we also map these height differences of shots to the site as Figure 4-6. It can be compared to Figure 4-3 and observed that 10 cm is a good threshold which helps to identify where the tree canopy regions are. Thus,

PAGE 35

35 the laser shots can be put into categories of multiple-return shots and single-return shots simply by comparing vertical distances of individual shots to 10 cm. Therefore, if the height difference of one shot is below 10 cm, it is categorized as a single-return sh ot. Otherwise, it is considered as a multiple-return shot because its first-return and last-return laser point s are reflected from two different spots in which we call potential tree spots and potential non-tree spots, respectively. Taking this city site as an example, we first ge nerate its DSM by resampling the irregular LiDAR point clouds to a regular grid image (Figure 4-7a). Af ter projecting the x and y coordinates of first-return points of multiple-return shots represented by black do ts to the corresponding DSM image, the fantastic characteristic of small-footprint LiDAR data can be observed again (Figure 4-7b). Those projected black dots in Figure 4-7b consist of tree tops, bushes, some edges of buildings and bridges, etc. The z coordinate of the first-return point is normally high er than the last-return point for any multiple-return shot. Hence, the firstand last-return points of mu ltiple-return shots are classified as the potential tree spots and potential non-tree spots, respectively (Figure 4-1). Morphological Filtering Two further challenges need to be conquered af ter potential tree/non-tree spots are identified: 1) how to reject false potential tree spots as the noncanopy cell, and 2) how to accept potential non-tree spots as the canopy cell (Figure 4-1). The canopy/non-ca nopy cell here is defined as a 1m by 1m gridded DSM image and determined by whether the cell is occluded by trees or not. False potential tree spots include edges or protrusions of buildings, edges of bridges, and any objects lower than normal trees. Potential non-tree spots could also be the actual tree s pots in the dense foliage case, because some height distances between leaves are probably less than the ve rtical accuracy of small-footprint LiDAR. These two challenges can be solved simultaneously by taki ng advantage of morphological filtering techniques. First, the binary image of tree spots is generate d by labeling every potential tree and non-tree cell as 1 or 0, respectively (Figure 4-8a). Then, we perform the morphological closing filter defined by [61] ) ( B A B A B (4.1)

PAGE 36

36 where B A and A B are the equation of morphological dilation and erosion, respectively. In this closing operation, the disk-shaped structuring element w ith a radius of 2 pixels is used to preserve the circular nature of the tree tops and recover missed tree spots occurred in the dense foliage case (Figure 4-8b). The closing filtering is followed by the morphological opening filter defined as [61] A B A( B B ), (4.2) where the same structuring element is used to rem ove false potential tree spots in Figure 4-8b which are sparser than actual tree spots. This filtering effect can be observed from Figure 4-8c. Since the opening will decrease the size of detected tree spots, mor phological dilating filtering of the same structuring element is performed to retrieve reduced edges of tree spots (Figure 4-8d). Because the dilation only increases the area where cell values ar e 1, the dilating filtering is able to prevent false alarm spots coming back since their cell values were changed from 1 to 0 by the morphological opening process. Tree Canopy Point Decision Based on previous morphological filtering procedur es, the vegetation area is detected and showed in Figure 4-9. Comparing the multiple return points ma rked in Figure 4-7b to the result of detected tree spots or canopy cell marked in Figure 4-9, it shows th at false potential tree spots are rejected and missed dense tree tops are recovered as well. However, the goa l of this tree-canopy removal design is to classify LiDAR points into canopy and non-canopy points (F igure 4-1). The canopy points here are defined as LiDAR points which are from the top part of trees and possess the ability to occlude under-canopy objects. Oppositely, the non-canopy points are any other points which are not tree canopy points. All the LiDAR points inside our non-canopy cells are classified as non-canopy points, but points at canopy cells are not always canopy points since both vegetation an d ground points probably exist at the same cell. Therefore, non-canopy points can exist in a canop y cell because some points inside the canopy cells could be reflected from the ground surface occluded by trees. In order to recognize the canopy/non-canopy points inside the same canopy cell, the ground reference is generated by resampling the irregular LiDAR points inside those non-canopy cells to a

PAGE 37

37 regular grid image. The above ground level (AGL) of each point then is easily obtained by calculating the height difference between its z coordinate and its corresponding gr ound reference. If the AGL of one point inside the canopy cell is greater than 0, it will be classified as a canopy point. Otherwise, it will be a non-canopy point (Figure 4-1). Tree Canopy Removal Result By removing those classified canopy points, the canopy removal result of the city test site (Figure 4-10) can be obtained from the remaining non-canopy points. When compared to its original DSM (Figure 4-7a), it is easy to see that most vegetation objects are successfully removed. This design also helps mitigate the vegetation impact in the bare-earth extraction algorithm. On the other hand, three forest test sites are i nvestigated to examine our canopy removal algorithm. First, The DSM of forest test sites (Hogtown parkin g site, Hogtown forest site and Hogtown residential site) are generated by resampling their irregular LiDAR point clouds to regular grid images showed as Figure 4-11a, Figure 4-12a, and Figure 4-13a, respec tively. After applying the canopy removal algorithm to these forest test sites, the new DSM without tree canopy can be obtained and are shown in Figure 4-11b, Figure 4-12b, and Figure 4-13b, respectively. By observing and comparing these forest test sites between their original DSM and canopy removal DSM, the difference is obvious that most forest canopy points have been successfully removed and all obscu re vehicles or buildings underneath canopy can be easily seen. In addition, since the canopy points have been cl assified from our algorithm, the occluded rate of forest canopy can be easily obtained by the following equation, % 100 Points # Points # Points # NonCanopy Canopy Canopy Rate Occluded. (4.3) Accordingly, the occluded rates for the Hogtown parking site, Hogtown fo rest site, and Hogtown residential site are calculated and they are 63.51%, 80 .59% and 48.31%, respectively. In fact, the original average LiDAR densities are 43.22, 50.13, and 36.72 points per square meter for the Hogtown parking site, Hogtown forest site, and H ogtown residential site, respectivel y. And, their remaining LiDAR

PAGE 38

38 densities after removing canopy points are 15.77, 9.73, and 18.98, respectively. Therefore, the vehicles underneath canopy in the Hogtown forest site will be more difficult to detect than in the two other sites. Furthermore, the detailed distribution of remaini ng point density can be found as well since all x-y locations of non-canopy points are known. The dens ity distributions for the Hogtown parking site, Hogtown forest site, and Hogtown residential site ar e shown in Figure 4-14a, Figure 4-14b, and Figure 4-14c, respectively. It is noted that this kind of di stribution is very useful for predicting the performance of occluded target detection with respect to various object locations. Summary This canopy removal algorithm is demonstrated wh ich helps 1) detect obscure targets underneath forest canopy and 2) mitigate the vegetation problem for those DTM extraction algorithms. As a matter of fact, the thermal imaging cameras can see the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fog, but not through the forest canopy. This proposed algorithm is learned from a city training site and verifi ed by two LiDAR systems, another city test site, and three forest test sites. Whether in a city or a fo rest site, the vegetation area can be correctly detected and canopy points are successfully removed. All obs cure vehicles or buildings underneath tree canopy are revealed as we demonstrated above. Accordingly, the occluded rate of forest canopy can be obtained. Furthermore, the detailed x-y distribution of remain ing point density can be found as well which will be very useful for predicting the performance of occlude d target detection with respect to various object locations.

PAGE 39

39 Figure 4-1. Tree-canopy removal design flowchart Figure 4-2. LiDAR multiple return shot illustra tion represented by black dots and overlapped on the training site.

PAGE 40

40 Figure 4-3. Snapshot image of the site22 from Google Earth. -5 0 5 10 15 20 25 30 35 0 0.5 1 1.5 2 2.5 3 x 104 shotwise point-based height difference shotwise height difference (m)freq Figure 4-4. Histogram of shot-wise height differences at the site22

PAGE 41

41 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 2000 4000 6000 8000 10000 12000 14000 shotwise point-based height difference shotwise height difference (m)freq Figure 4-5. Zoom in the smaller difference part of Figure 4-4 and set the clipping range from 0 to 1. x (meter)y (meter) 20 40 60 80 100 120 20 40 60 80 100 120 0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0 1 Figure 4-6. Shot-wise height difference map clipped to 0.0m-0.1m at the site22.

PAGE 42

42 A B Figure 4-7. Multiple return map at the samp22 city site: A) Original DSM. B) Multiple return points represented by marked circles. A B C D Figure 4-8. Morphological filtering results at the samp22 city site: A) Binary image of Canopy/ Non-Canopy cell represented by white/black color at the city test site. B) Image of applying morphological closing filter to Figur e 4-8a. C) Image of applying morphological opening filter to Figure 4-8b. D) Image of applying morphological dilating filter to Figure 4-8c.

PAGE 43

43 Figure 4-9. Detected vegetation area (canopy ce ll) marked on DSM of the city test site Figure 4-10. Canopy removal result for the city test site A B Figure 4-11. Revealed occluded vehicles at the H ogtown parking site: A) Original DSM B) Canopy removal DSM where vehicles are circled by wh ite lines in which 7 occluded vehicles are revealed.

PAGE 44

44 A B Figure 4-12. Revealed occluded vehicles at the Hogtown forest site: A) Original DSM B) Canopy Removal DSM where the occluded tarpaulin, mi mic vehicle, circled by white lines is revealed. A B Figure 4-13. Revealed occluded vehicles at the Hogt own residential site: A) Original DSM B) Canopy Removal DSM where vehicles are circled by wh ite lines in which 4 occluded vehicles are revealed. A B C Figure 4-14 Remained point density in 2-D distri bution after removing canopy points in A) Hogtown parking site, B) Hogtown forest site and C) Hogtown residential site.

PAGE 45

45CHAPTER 5 POINT-BASED BARE-EARTH EXTRACTION DESIGN In the last decade, a variety of algorithms have been developed for extracting the DTM from point clouds generated by LiDAR systems, where automa tic and robust ground point extraction has been attracting great attention. Although most filters perform well in flat and uncomplicated landscapes, landscapes containing steep slopes and discontinuiti es are still problematic. We developed a novel bare-earth extraction algorithm and compared our pe rformance to a dozen filters working on the same fifteen study sites which were provided by ISPRS. Th e result is that the average total error and kappa index of agreement of our algorithm in the automa ted process is 4.6% and 84.5%, respectively, which both outperform all other twelve pr oposed filters. Our kappa index, 84. 5%, can be interpreted as almost perfect agreement. Segmentation Modeling The proposed bare-earth extraction design, consis ting of segmentation modeling (Figure 5-1) and surface modeling, is the second design part of this entire work. The input data, remaining non-canopy points, of our segmentation modeling are obtained by removing those canopy points from our tree-canopy re moval design. This segmentation modeling is a triangulated irregular network (TIN) based design in cluding triangle assimilation, edge clustering, and point classification and aims to achieve better discrimination of objects and preserve terrain discontinuities in the segmentation modeling. The triangle type, which is either flat or steep, is assimilated and reclassified by iteratively calling the ma jority vote from its neighboring triangles. In the surface modeling, the point is reclassified by using the roughness estimation of canopy removal DSM and DTM, bridge detection, and sharp ridge detectio n to further reduce both Type I and Type II errors. An extra advantage of this point-based algorithm is that we can avoid choosing any filtering window size during a pixel-based processing. Triangle Classification and Assimilation At the beginning, the Delaunay tria ngulation is used to build up the Triangulated Irregular Network

PAGE 46

46 (TIN). Two properties make the Delaunay triangulation attractive [62]. First, it is a local definition for triangulation. The insertion of a new point need only ex amine that region in the plane which is closer to the new point than any other point. Second, it is known that a given existing triangle is affected only if a new point is added within the circle circumscribed through its vertices. This prescribes a well-defined local search area for new points which can improve the fit of the terrain model to the region inside the triangle. The TIN is a network of non-overlapping triangl es generated from irregularly distributed LiDAR points with three dimensional coordinates (x, y, a nd z). Based on the TIN, each triangle with its three points can be classified either as a flat or steep tr iangle by calculating the height difference between its highest and lowest point and the slope degree, which is the degree difference between normal direction to the triangle plane and the direction to the sky parallel to the z axis (Figure 5-2a). If its height difference, h is lower than 0.3 meter and slope degree, is smaller than 45 this triangle will be classified as a flat triangle. Otherwise, it will be classified as a steep triangle. The threshold for height difference, h should be as small as possible to ensure the quality of the flat triangles. Since the vertical accuracy of the LiDAR data is 0.3 meter, it is a good choi ce for the height difference threshold. Although all triangles can be categorized into either flat or steep triangles, this simple decision rule of comparing the height difference and slope degree ca nnot give a satisfactory categorization. Hence, we add on the following novel triangle assimilation to get improved triangle classification. Every triangle has three edges which are shared with at most, three other triangles. From those 3 triangles, we are able to give a better triangle classification, steep or flat, by taking a majority vote. For instance, there are three bold edges of the triangle A which are shared by triangles B, C, and D in Figure 5-2b. In this triangle group, if there are two or more of the adjacent triangles (B, C, and D) belonging to the flat triangle class, then triangle A will be classified as a flat triangle, wh ere the vote of flat to steep could be 3-0 or 2-1. Likewise, if the number of adjacent triangles B, C, and D classified as steep triangles is more than flat triangles, then the triangle A will be assimilated to the steep triangle class, where the vote of flat to steep could be 0-3 or 1-2. During this assimilation process, the higher agreement vote 3-0 or 0-3 has to have

PAGE 47

47 the higher priority than the lower agreement vote 2-1 or 1-2. Therefore, those triangle groups giving higher agreement vote are searched first and used to assimilate their central triangle. Then, the other triangle groups having lower agreement votes are followed for the triangle assimilation. The above triangle assimilation composed of two ki nds of agreement votes has to be iterated until the number of triangles needed to be assimilated can not be decreased any more. Figure 5-2c shows the fewest number configuration for a stable triangle clus ter, which consists of uniformly classified, either flat or steep, and edge-connected triangles E, F, and G. Any clusters including less than three triangles or in unstable configurations continue to be reclassifi ed during this iterated triangle assimilation. Hence, the triangle classification will become better and better th rough step by step iterations since the number of triangles which need to be assimilated becomes less and less (Figure 5-3). The number of reclassified or assimilation-needed triangles can be regarded as the desirable noise level to be removed. Therefore, the performance of this triangle classification can be represented as the reduced noise in dB by the following formula, Process NoiseAfter eProcess NoiseBefor 10 log 10 se ReducedNoi. (5.1) For this case in the samp22 site, the reduced noise is equal to 96 / 2572 log 1010 =14.3dB. This value shows how meaningful the iterated triangle assi milation is. The performance is contributed to the fact that not only are three immediate neighbor tr iangles considered to make a decision whether the center triangle is a steep or flat triangle, but the iteration steps also extend the assimilation process to the global area by the iterative procedure, which are like chained reactions. The triangle assimilation has the capability of performing a local search while preserving the merits of global treatment. Therefore, the final stable assimilation gives a triangle classifica tion result which is from the majority vote among neighboring triangles locally and globally. Edge Classification and Clustering Every triangle is composed of 3 edges, each of which can be shared by, at most, two adjacent triangles. The combination cases of these two adjacen t triangles can be both flat, both steep, or mixed

PAGE 48

48 which is one flat and one steep triangle. For these th ree cases, each edge can be classified into three categories: flat, steep, and boundary, respectively. For example suppose that Figure 5-4 shows a small part of the TIN and is processed by the above tria ngle classification. There are three shaded and three white triangles which represent flat and steep triangl es, respectively. Then, th e dash lines, bold lines (A, B, and C), and regular lines are going to be classified as the flat edges, boundary edges, and steep edges, respectively, based on the above rule. Although the regul ar lines or steep edges are not shared with other adjacent steep triangles in Figure 5-4, they must be in reality. These triangles are not shown here in order to keep the figure concise. Otherwise, these three white, or steep, triangles should be assimilated to be flat triangles by the majority vote. One object could include many edges of the TIN. Th e edge clustering then is needed to represent objects and three kinds of edge clusters are formulat ed: flat edge clusters, boundary edge clusters, and steep edge clusters. The edge clustering is simply used to cluster those edges which are connected together and have the same edge class. Figure 5-5 sh ows the edge classification result of the samp22 site where the flat and steep edges are represented by black and gray lines, respectively. Since the TIN here is a vector based representation of the physical surfa ce built up by the nodes and lines from LiDAR point clouds, our edge classification is vector based. Fi gure 5-6 is the boundary edge classification result which is cluster-wise and indicates individual bounda ries for all potential objects. The advantage of our boundary edge classification is that the spotty and disconnected edges, which are of most concern in edge detection methods of image processing, do not exist in our cluster-wise and vector-based edge detection (Figure 5-6). Point Classification While all points inside a steep edge cluster obt ained from the above edge classification can be simply classified as object points, classifying points inside a flat edge cluster is a challenge since a flat edge cluster could be either a gr ound or roof surface. This challenge can be solved by calculating the ground ratio for each flat edge cluster:

PAGE 49

49 t) SteepHeigh latHeight Boundary(F # t) SteepHeigh latHeight Boundary(F # o GroundRati, (5.2) where the denominator is the number of boundary edges where flat triangles are higher than steep triangles, both of which have to share the same boundary edges, and the numerator is just to calculate the other lower part. Taking Figure 5-4 as an example again, a flat edge cluster is formed by three dashed lines which include the point P0 and three white points. Its boundary edges are three bold lines A, B, and C. The height comparison between a flat and steep tr iangle sharing the same boundary edge A, B, or C is equivalent to calculating the height difference betw een one of the black points P1, P2, or P3 and the point P0. Suppose that the point P0 is lower than other points P1, P2, and P3. Then, the denominator in Equation (5.2) is equal to 0 and the numerator is 3. Th e ground ratio of this flat edge cluster is equal to 3/0, or infinity. Conversely, if the point P0 is higher than points P1, P2, and P3, then this cluster-wise ground ratio is equal to 0/3, or 0. The ground ratio is a good indicator for how confiden tly we classify a flat edge cluster to a ground surface. In general, when the ground ratio is greater th an 1, this flat edge clus ter has higher probability to be ground than roof. However, when many boundary edges of the roof are connected to other higher parts of buildings, this ground ratio could be greater than 1. Then, the threshold for ground ratio should be considered as strict as possible to ensure the quality of extracted ground points. Eventually, the threshold for ground ratio is set at 8 by our experimental tests, where the error of accepting roof points as ground points is minimized. Therefore, all points in side a flat cluster would be classified as ground points if its ground ratio is higher than our threshol d. Otherwise, those points would be classified as object points. In addition, all points inside steep clusters are classified as object points simply because those points do not belong to flat clusters. Surface Modeling The above segmentation modeling is to ensure th e quality of ground points such that those extracted ground points can safely be considered as actual ground points and act as seed points to progressively extend the ground regions. Our extract ed ground/object points can be compared to

PAGE 50

50 reference ground/object points by the cross matrix in Table 5-1, where the False Negative (FN) case, or Type I error, is for falsely rejecting ground points, an d the False Positive (FP) case, or Type II error, is for incorrectly accepting object points. The remaining True Positive (TP) and True Negative (TN) cases are for the correct classifications of ground and object points, respectively. Figure 5-7 shows the result after applying the me thod in the previous section to the samp22 site, where the Type I error is represented by black x-ma rks and the Type II error is represented by white x-marks. It is noted that the Type I error is very high and Type II error is very low in Figure 5-7. That is by design of our bare-earth extraction technique becau se the ratio of TP to FP will be high due to low Type II error so that all extracted ground points should be actual ground points. In order to reduce both Type I and Type II errors, the four-phase point reclassification based on the surface modeling is developed and its flowchart is shown as Figure 5-8 where each phase is iterated. Phase-I Point Classification Based on DTM In the phase I reclassification, the terrain raster map is generated by resampling the extracted ground points to a regular grid image and used as th e ground level reference, where the grid cell size is 2 meters by 2 meters for all 15 sites. Due to the go od quality of extracted ground points in which Type II error is very low, it is possible to recover some mi ssed ground points by comparing their height levels to the terrain raster map. To do so, the AGL of each point is calculated by s ubtracting its corresponding elevation on the terrain raster map from its z coordina te. If the AGL of one ground point is higher than the height threshold defined as the variable TH1, it is r eclassified as an object point. On the other hand, if the AGL of one object point is lower than the TH1, it is reclassified as a ground point. The choice of TH1 will influence the result of th is phase reclassification. With our experiments, setting the TH1 as 1.6 meters is a good choice for all 15 sites. However, it can be adjusted manually to obtain optimal results for each site. In each iteration, those updated ground points will be used to update the terrain raster map as the ground level reference fo r the next iteration. The total number of corrected ground and object points will be decreased step by step with each iteration. The iterative process ends when the total corrected number cannot be decreased any more. The result of the phase I reclassification

PAGE 51

51 in the samp22 site is shown as Figure 5-9, where 1.6 meters is adopted for TH1. Comparing Figure 5-7 to Figure 5-9, the Type I error is obviously reduced whic h is the goal of the reclassification in this phase I. Phase-II Point Classification Based on DSM Although the Type I error is reduced in the phase I reclassification, the Type II error is increased when comparing Figure 5-7 to Figur e 5-9. Thus, this phase II reclassification is designed to reduce the Type II error resulting in the phase I reclassification. First, the roughness raster map is generated from the previous non-canopy points because tree canopy points will impact the estimation of surface roughness. The chosen pixel size and window size is 0. 8 meters and 4 meters, respectively, for each side. In each running window, its roughness value here is defined as the elevation difference between the highest and lowest pixel among 25 pixels (5 pixel by 5 pixel square) located in this window. After the roughness map is generated (Figure 5-10), all grou nd or object points can find their own corresponding roughness values by their x and y coordinates. The ra ster DTM has to be generated by resampling the current ground points for the ground level reference of a ll points, where the grid cell size is 0.8 meters by 0.8 meters for all 15 sites, and then the AGL values of all points can be calculated by their z coordinates and the obtained DTM. Based on obtained roughness and AGL values, each point can be reclassified by the following rule. If one object point has a roughness value less than 0.3 meters and its AGL is lower than 0.3 meters, then this point will be reclassified as a ground point. Since a flat local area, 5 pixels by 5 pixels, normally consists of either all ground points or all object points, a point inside the area and close to the DTM should be a ground point rather than an object point. Similarly, if one ground point has a roughness value greater than 0.3 meters and its AGL is higher than 0.3 meters, then this point will be reclassified as an object point. Since a rough local ar ea usually consists of bot h object and ground points, a point inside the area and above the DTM should have a higher probability of being an object point than a ground point. In each iteration of this phase, updated ground points will be used to update the DTM in the next iteration. The roughness map does not have to be updated since it is generated from all points. The total number of corrected ground and object points will also be decreased step by step with each iteration. The

PAGE 52

52 iterative process ends when the total corrected number cannot be decreased any more. Figure 5-11 shows the result of phase-II point reclassification in the sam p22 site. By observing Figure 5-9 and Figure 5-11, it is obvious that the Type II error is dramatically reduced by this phase reclassification. It is the roughness map that helps to reduce the Type II error because most false ground points which are corrected to object points are located at high roughness areas, which can be found by mapping those white x-marks of Figure 5-9 to the corresponding roughness map of Figure 5-10. Phase-III Point Classification Based on DTM Roughness The sharp ridges on the ground surface are usually a major problem for ground filtering [15]. This problem can be mitigated by this phase-III point reclassification. First, a roughness map for DTM (Figure 5-12) is generated by resampling those grou nd points from phase-II point reclassification, where the chosen pixel size and window size are the same as before. Second, we propose a simple and effective bridge detection for the later correction. Our bridge detection can be separated into two part s: bridge body and bridge ends. For the bridge body, the morphological top-hat filtering with a disk-s haped structuring element and a radius of 12 pixels is performed on the canopy removal DSM, where the top-hat is useful for enhancing detail in the presence of local tops [61]. Then, this top-hat transf ormation is converted to a binary image where pixel levels larger than 4 are changed to 1 and other pixels are changed to 0. This is followed by the opening transformation with a disk-shaped structuring element a nd a radius of 2 pixels to remove small isolated noise. For bridge ends detection, the roughness of DT M (Figure 5-12) is useful since the bridge ends always occur at sharp ridges on the ground surface. Then the other binary image for possible bridge ends is generated where those roughness values of pixels gr eater than 0.7 meters are changed to 1 and other pixels are changed to 0. By overlapping these tw o binary images, any one bridge body which is connected to two bridge ends is our final detected bridge. The LiDAR points of detected bridges are represented by black x-marks in Figure 5-13 which show no obvious missing or false alarm errors in our bridge detection. Based on the roughness map of DTM, it is easy to know where the regions of sharp ridges on the

PAGE 53

53 ground surface are. For those problematic areas, each rough area of DTM can be corrected by its nearest flat part in which the height level of the flat part will extend to the rough part. This approach can shrink the width of rough areas since those rough parts have been updated to the same ground levels of their nearest neighboring flat parts of the DTM. Therefore, those object points in rough regions of the DTM are reclassified as ground points if their AGL from the updated DTM are lower than 0.3 meters. It is noted that the detected bridges have potential to be reclassified as ground points, so all bridge points are detected first and are excluded in this rule. In add ition, those ground points in flat regions of the DTM are reclassified as object points if their AGL from the updated DTM are higher than 0.3 meters. In each iteration of this phase, the DTM is upda ted to make the width of rough areas become narrower and narrower. The iteration ends as before when the total number of corrected ground and object points cannot be smaller than previous iteratio ns. The result (Figure 5-14) shows that the number of Type I errors occurring in the sharp ridges on the ground surface is indeed decreased significantly which can be easily seen by checking the road sides n ear the bridge in the left bottom parts of Figure 5-11 and Figure 5-14. Phase-IV Point Classification Based on Flattened DSM It is easy for the final point reclassification to achieve good performance when most of the sharp ridge points are able to be extracted as ground points. First, the DTM is generated by resampling those ground points from phase-III point reclassification, wher e the grid cell size is 0.7 meters by 0.7 meters for all 15 sites. Then, the AGL values of all points can be obtained by referring to the generated DTM in this phase. The flattened DSM (Figure 5-15) is ge nerated by resampling all points with AGL values instead of their original z coordinates, where it is clear to see that most of the rough areas including sharp ridges on the ground surface are flattened, especially fo r those road sides near the bridge in the bottom left part of Figure 5-15. The ground/object points can simply be reclassifi ed from the flattened DSM by their AGL values obtained in this phase. If object points are lower than the threshold TH2, those points are reclassified as ground points. Conversely, if ground points are higher than the threshold TH2, those points are

PAGE 54

54 reclassified as object points. This also needs itera ting to make the flattened DSM become flatter and flatter in those rough terrain areas. The iterated pr ocess ends, as before, when the total number of corrected ground and object points of the current iteration is not reduced from the previous iteration. The TH2 here is applied as 0.6 meters for all 15 sites. It can be manually adjusted to reach a better performance. The Type I and Type II errors of th e final classification of ground/object points are represented by black and white x-marks, respectivel y, in which the TH1 and TH2 are applied by 1.6 meters and 0.6 meters, respectively (Figure 5-16). Bare-Earth Extraction Result The following two sections show the automatic and manual performances of our proposed bare-Earth extraction design, respectively. In orde r to evaluate our algorithm, the open data sets from ISPRS were used in order to compare our results to other algorithms using the same data sets. Same Parameters for All 15 Sites of ISPRS Through processes of our tree-canopy removal, bare -earth extraction, and point reclassifications, the final point classification results for all 15 study si tes are shown, where the Type I and Type II errors are represented by black and white x-marks, r espectively, and marked on their corresponding canopy removal DSM images. In order to show the geographi c relation between the study sites and their original sites, the corresponding areas are also marked by thei r boundaries (Figure 5-17 ~ Figure 5-38). It is noted that both applied parameters TH1 a nd TH2 are fixed and the same for all 15 sites, which are 1.6 meters and 0.6 meters, respectively. The accuracy of the classification map is assessed w ith the use of the error matrix from which the total error and the Kappa index of agreement are deri ved [64]. The total error is calculated by adding up the Type I and Type II errors and dividing this sum by the total number of points, so this is the overall probability of a reference pixel being incorrectly classified. The kappa index of agreement is an alternative measure of the overall classification accuracy that subtracts the effect of chance agreement and quantifies how much better a particular classifica tion is, as compared to a random classification [65]. The equation for the kappa index of agreement [66], is

PAGE 55

55 e e aPr 1 Pr Pr (5.3) where Pr( a ) is the relative observed agreement among raters, and Pr(e) is the hypothetical probability of chance agreement, using the observed data to cal culate the probabilities of each observer randomly deciding each category. Accordingly, the Pr(a) and Pr(e) can be derived to the following equations TN FP FN TP TN TP a Pr (5.4) 2) ( ) )( ( ) )( ( Pr TN FP FN TP TN FN TN FP FP TP FN TP e (5.5) where the TP, FN, FP, and TN are defined as in Table 5-1. The significance of the kappa index of agreement is interpreted by Table 5-2 [47]. The accuracy of our point classification for every st udy site is assessed by the percentages of total errors and the kappa index of agreem ent (Table 5-3). The dense objects on steep terrain in a city site will cause more false positive, or Type II, errors (accept obj ect points) which occur in the site of samp11. The rough and narrow terrain with sharp ridges will cause more false negativ e, or Type I, errors (reject ground points) which happens in the sites of samp53 and samp61. Other difficult terrain types, however, do not im pact our performance. For example, the complex scene in the site of samp23 is a plaza surrounded on three sides by a block of buildings. There is a sunken arcade in the center of the plaza. Both the pl aza and arcade should be assumed to be bare-Earth. For this scene, the filters that make use of local su rface assumptions like in Axelsson, Preifer, and Shon performed best [42]. In our filter, the ground ratio for judging each flat edge cluster in the TIN is also a rule based on the local surface in which the height difference of its all connected clusters will be considered rather than comparing its height to the lo cal lowest place. Thus, this complex scene is not a problem for our filter. Bridges can also be a difficult terrain type. The Type II error (classify object points as ground points) usually occurs near the beginning and end of bridges while bridges in the test should be treated as objects

PAGE 56

56 [42]. In our filter, the bridge is detected by searching for and connecting its bridge body and both beginning and ends. The detected bridges including both ends are treated as object points instead of ground points such that this kind of Type II error will be reduced. Similar to bridges are ramps. Ramps bear similar ity to bridges in that they span gaps in the bare-Earth. However, they differ in that they do not allow movement below them. As such, ramps were treated as bare-Earth in the reference data. All th e tested algorithms filtered off the ramps [42] and caused Type I errors. In our filter, most points on ramps are initially classified as object points. By iterating point reclassification, those points initially cl assified as object points but near to bare-Earth can be corrected and reclassified as ground points. Thus, even though only partial and lower points on ramps are classified as ground points, they will grow upward and extend gradually to cover most of the ramps through iterated correction. Accordingly, those ramps ar e treated as the bare-Earth in our filter. It can be checked in our point classification result for the site of samp71, where ramps are used to connect the road across the bridge. Vegetation on slopes can be another difficult terrain type [42], but this is not an issue for our filter. At the beginning, most vegetation points are already re moved in our canopy removal design where the fact that vegetation objects usually reflect di fferent height levels for the first and last return of each laser shot is used. Besides, it is noted that vegetation will im pact the roughness estimation for bare-Earth. As such the canopy removal design becomes very meaningful in vegetated terrain. For the accuracy assessment, our filte r in the same parameters case or automated process (called as Chang#1 here) is compared to the best two filters in the report [68]. The total errors and kappa index of agreement for each site are represented by empty a nd filled legends, respectiv ely (Figure 5-39). This result shows that Chang#1’s filter often gives better performance than Pfeifer’s filter and is little better than Axelsson’s filter. The average total errors and kappa index of agreement for 15 sites are compared and given in the next subsection. Optimized Parameters for All 15 Sites of ISPRS By testing different associated parameters of our filter in all study sites and comparing their

PAGE 57

57 influences, it is found that the TH1 and TH2 values will dominate the final performance. Accordingly, our filter is optimized by adjusting these two paramete rs to substitute the initial values of TH1 (1.6m) and TH2 (0.6m). The accuracy of our point classifi cation for every study site with two optimized parameters, TH1 and TH2, is liste d in Table 5-4, which also includes corresponding FN and FP. It is found that the smaller the TH1 is, the better the abilit y to detect the smooth terrain will be. Conversely, the larger the TH1 is, the better the ability to mitig ate the sharp ridges will be. On the other hand, the TH2 is applied to try to correct those errors from th e TH1. Hence, the more the error from the TH1 is, the larger the TH2 will be. This phenomenon can be observe d in the TH2 column of Table 5-4 for the sites samp11, samp52, samp53, and samp61. The automated (Chang#1) and optimized (Chang#2) r esult is also compared in Figure 5-40. It shows that Chang#2’s filter always performs better th an Chang#1’s filter. Figure 5-41 shows the average performance of all study sites for those filters in [6 8] and both of our automated and optimized filters, where the average total errors and average kappa inde x of agreement are calculated for each filter. It is clear that our automated filter outperforms all other filte rs in [68] and the performance is even better after two dominant parameters are optimized manually. Recently, Sithole [26], Silvan-Cardenas [27], Lu [28], and Meng [29] proposed new filtering algorithms and used the same data sets to assess thei r performance. Some authors presented their filtering results by the kappa index of agreement only, so we have to calculate other authors’ TP, FN, FP, and TN values defined in Table 5-1 of each site to obtain their average kappa value for 15 study sites for comparing the performance difference among filters. Th e average accuracy compar ison of 15 study sites for our automated and optimized algorithms to four di fferent filters proposed recently (after year 2004) is represented by the kappa index of ag reement and shown in Figure 5-42. It is noted that these recent filtering performances are even worse than the filter developed by Axelsson in 2000. However, both of our automated and optimized filters are better than a ll these twelve filters, in which eight were proposed before 2004 and four were proposed after 2004.

PAGE 58

58Summary Automatic and robust ground filtering is important in LiDAR applications where classified ground and object points can be used for DTM generation and further reconstruction of topographic features. Many methods have been proposed to extract point s on bare-Earth from LiDAR data. However, most filters perform well in flat and uncomplicated la ndscapes, while the landscapes containing steep slopes and discontinuities are still a problem which has not been fully solved. We present an algorithm which is composed of segmentation modeling and surface modeling. The special features inside our designs include the TIN based platform, iterated triangle assimilation, edge clustering, roughness estimation of canopy removal DSM and DTM, bri dge detection, and sharp ridge detection, etc. These designs aim to mitigate vege tation interference, obtain better discrimination of objects, preserve terrain discontinuities, and reduce both Type I and Type II errors. The performance of our algorithm is compared to twelve proposed filters and evaluated by working on the same fifteen study sites. The average total errors and kappa index of agreement of this work in the automated process is 4.6% and 84.5%, respectively, wh ich outperforms all twelve other filters and such kappa index is interpreted as almost perfect agreement. In addition, this work applied with optimized parameters performs even better.

PAGE 59

59 Table 5-1. Extracted and reference cross matrix Points Extracted Ground Extracted Object Reference ground True positive (TP) False negative (FN)/Type I error Reference object False positive (FP)/Type II error True negative (TN) Table 5-2. Interpretation of the kappa index of agreement Kappa index Interpretation < 0% No agreement 0% ~ 20% Slight agreement 21% ~ 40% Fair agreement 41% ~ 60% Moderate agreement 61% ~ 80% Substantial agreement 81% ~ 100% Almost perfect agreement Table 5-3. Total errors and kappa index for 15 st udy sites with same parameters (TH1=1.6m and TH2=0.6m) in our filter. Site TP TN FNFPErrorKappa Samp11 20421 11943 1365428114.9%68.9% Samp12 26292 23534 39918944.4%91.2% Samp21 10081 2675 42001.6%95.3% Samp22 22299 9159 20510433.8%90.9% Samp23 12919 11169 3047034.0%91.9% Samp24 5161 1846 2732126.5%83.9% Samp31 15522 12393 349133.3%93.4% Samp41 5408 4981 1946487.5%85.0% Samp42 12377 29199 668282.1%95.0% Samp51 13915 3250 356453.8%88.2% Samp52 19679 1857 4335054.2%77.5% Samp53 31156 1217 18331725.8%52.2% Samp54 3890 4441 931843.2%93.5% Samp61 33354 991 5002152.0%72.4% Samp71 13781 1495 942752.4%87.7% Table 5-4. Total errors and kappa index of agreement for 15 study sites in our filter with two optimized parameters, TH1 and TH2. Site TP TN FNFPErrorKappa Samp11 0.5 m 0.8 m 1158390313.3%72.2% Samp12 1.0 m 0.5 m 57112623.5%93.0% Samp21 0.5 m 0.5 m 241351.2%96.4% Samp22 1.1 m 0.5 m 2695742.6%94.0% Samp23 1.5 m 0.6 m 2987074%92.0% Samp24 0.6 m 0.5 m 2791976.4%84.3% Samp31 0.5 m 0.4 m 992561.2%97.5% Samp41 2.4 m 0.5 m 1856107.1%85.9% Samp42 0.9 m 0.5 m 1084631.3%96.8% Samp51 0.3 m 0.5 m 774803.1%90.5% Samp52 2.2 m 0.9 m 1516343.5%79.6% Samp53 2.6 m 1.1 m 6474653.2%60.8% Samp54 1.2 m 0.7 m 801622.8%94.4% Samp61 2.4 m 0.8 m 1532341.1%82.8% Samp71 1.0 m 0.7 m 792442.1%89.3%

PAGE 60

60 Figure 5-1. Segmentation modeling flowchar t of our bare-earth extraction design Figure 5-2. Triangle classification and assimilation illu stration: A) The height difference h and slope degree of a triangle. B) Triangle assimilation illustration. C) The fewest number configuration for a stable triangle cluster. Figure 5-3. Triangle assimilation curve of samp22 site

PAGE 61

61 Figure 5-4. Cluster-wise ground ratio illustration. Figure 5-5. Vector-based edge classification where bl ack and gray lines represent flat and steep edges, respectively. Figure 5-6. Cluster-wise boundary edge detection

PAGE 62

62 Figure 5-7. Initial point classification result where T ype I (black x-mark) and Type II (white x-mark) errors are marked on canopy removal DSM image. Figure 5-8. Surface modeling flowchart of our bare-earth extraction design

PAGE 63

63 Figure 5-9. Phase-I point reclassification result: Type I (black x-mark) and Type II (white x-mark) errors marked on canopy removal DSM image. Figure 5-10. Roughness map of canopy removal DSM Figure 5-11. Phase-II Point reclassification result: Type I (black x-mark) and Type II (white x-mark) errors marked on canopy removal DSM image.

PAGE 64

64 Figure 5-12. Roughness map of DTM Figure 5-13. Bridge detection result showed by ma rking detected bridges as black x-marks on the canopy removal DSM. Figure 5-14. Point reclassification phase-III result: Type I (black x-mark) and Type II (white x-mark) errors marked on canopy removal DSM image.

PAGE 65

65 Figure 5-15. The flattened DSM map obtained by resampling all points where their z coordinates are replaced by the above ground le vels referred to those ground points extracted from the phase-III point reclassification. Figure 5-16. Final point classification result: Type I (b lack x-mark) and Type II (white x-mark) errors marked on canopy removal DSM image. Figure 5-17. DSM image of the city site #1, where the marked rectangular a and b is the boundary of the study site samp11 and samp12, respectively.

PAGE 66

66 Figure 5-18. The final classification errors in the site of samp11. Figure 5-19. The final classification errors in the site of samp12. Figure 5-20. DSM image of the city site #2, where the marked rectangular a, b, c, and d is the boundary of the study site samp21, samp22, samp23, and samp24, respectively.

PAGE 67

67 Figure 5-21. The final classification errors in the site of samp21. Figure 5-22. The final classification errors in the site of samp22. Figure 5-23. The final classification errors in the site of samp23.

PAGE 68

68 Figure 5-24. The final classification errors in the site of samp24. Figure 5-25. DSM image of the city site #3, where the marked rectangle is the boundary of the study site samp31. Figure 5-26. The final classification errors in the site of samp31.

PAGE 69

69 Figure 5-27. DSM image of the city site #4, where the marked rectangles a and b are the boundaries of the study sites samp41, and samp42, respectively. Figure 5-28. The final classification errors in the site of samp41. Figure 5-29. The final classification errors in the site of samp42.

PAGE 70

70 Figure 5-30. DSM image of the forest site #5, wh ere the marked rectangles a, b, c and d are the boundaries of the study sites samp51, samp52, samp53, and samp54, respectively. Figure 5-31. The final classification errors in the site of samp51. Figure 5-32. The final classification errors in the site of samp52.

PAGE 71

71 Figure 5-33. The final classification errors in the site of samp53. Figure 5-34. The final classification errors in the site of samp54. Figure 5-35. DSM image of the forest site #6, where the marked rectangle is the boundary of the study site samp61.

PAGE 72

72 Figure 5-36. The final classification errors in the site of samp61 Figure 5-37. DSM image of the forest site #7, where the marked rectangle is the boundary of the study site samp71. Figure 5-38. The final classification errors in the site of samp71

PAGE 73

73 Figure 5-39. Accuracy comparison of our automated f ilter (Chang#1) to the best two filters (Pfeifer and Axelsson) in [68] for all 15 study sites. Figure 5-40. Accuracy comparison of our automate d filter (Chang#1) to our optimized filter (Chang#2) for all 15 study sites.

PAGE 74

74 Figure 5-41. Average accuracy comparison of 15 study sites to eight different filters proposed before year 2004 represented by the total e rror and kappa index of agreement. Figure 5-42. Average accuracy comparison of 15 study sites to four different filters proposed after year 2004 represented by the kappa index of agreement.

PAGE 75

75CHAPTER 6 GRID-BASED BARE-EARTH EXTRACTION DESIGN Object Detection Outlier and Tree Canopy Removal Local LiDAR point outliers are often randomly distributed over a study area which may be caused by passing birds, airplanes, or the sensor itself. Normally they are extremely higher or lower than adjacent points and isolated from other points whic h must be removed during preprocessing [29,69-72]. The simplest way to identify these outliers is to ex amine the frequency distribution of elevation values [29,72,73]. Manual examination of the da taset is another viable option [72]. In this study, only the last return points of LiDAR data need to be considered since none of the bare-Earth points come from the other return points. In the last return data set, a 5m 5m window iw is used to get neighbored LiDAR points for each point ip where its z coordinate is iz. The point ip is decided as an outlier if i i iz 2 (6.1) where the mean (i ) and standard deviation (i ) are obtained from those points inside the window iw. It is assumed that the height values belong to a normal distribution in which values less than one standard deviation from the mean (dark blue) account for abou t 68% of the set, while two standard deviations from the mean (medium and dark blue) account for ab out 95%, and three standard deviations (light, medium, and dark blue) account for about 99.7% (Fig ure 6-1). So, Equation (6.1) results in the 95% confidence intervals. As an example, in site22 there are 69 outlier points found by Equation (6.1) in all 32,707 last return points which occupy 0.21% of total points shown as Figure 6-2, where 20 points are lower and 49 points are higher than two standard deviations. After we get outlier-free data sets, the developed tree canopy removal algorithm in Chapter 4 is used again to ge t canopy-free sites, which are ready for the following segmentation and classification in this object detection section.

PAGE 76

76Segmentation by Non-Flat Regions In LiDAR data, the ground points are the measuremen ts from bare-Earth terrain that are usually the lowest surface features in a local area. Non-ground points are the measurements from the objects above the bare-Earth terrain, such as trees, buildings, bri dges, and shrubs. Surface slope is generally lower between two neighboring bare ground points than between one bare ground and one non-ground point [74]. Those non-ground points between objects and ba re-Earth points usually generate the non-flat regions which can be easily observed by watching the slope map of a DSM. Thus, the slope feature is used in this study and the 3rd order finite difference (3FD) algorithm is selected and defined by 2 2arctany xf f S (6.2) g z z z z z z fx67 9 4 6 1 3 (6.3) g z z z z z z fy63 9 2 8 1 7 (6.4) where g is spatial resolution (i.e., grid cell size) set as 1m and 0.5m for ISPRS and UF ALSM study sites, respectively, and the associated positions of z coordinates are assigned by Table 6-1. The 3FD derived slope method was recommended by [75] since it is less sensitive to the LiDAR data error which makes it more appropriate for applications. In addition, a low-pass 3 3 filter (Table 6-2) is applied to the derived slope map to reduce furthe r noise and its kernel weights are defined as 1/16 of values shown in Table 6-2. With smoothed slope inform ation, the surface steepness is easily detected by morphlogical tophat filtering which returns the imag e minus the morphological opening of the image (erosion followed by dilation). Based on our experiments, the disk structure element with radius 2 is a good choice for the tophat transform. Similarly, a smoothed tophat transform is obtained with the smoothing low-pass filtering. A pixel is decided to be flat if its smoothed slope is less than 45 and smoothed tophat value less than 8. Then, a flat re gion is built up simply by clustering those connected flat pixels. Figure 6-3 shows as an example for th e smoothed slope map and th e smoothed tophat filtering at the site22. Figure 6-4 shows its flat region decided by our rule.

PAGE 77

77Classification of flat regions A flat region could be a ground or an object region such as a building roof, car roof, etc. In order to classify these flat regions into either a ground or non-ground class, we extracted four features and put them into the rule-based methods to do this classi fication task. These features are: 1) the number of inside clusters, 2) height indicator, 3) area size, and 4) slope indicator of a tested flat region. Feature #1, the number of inside clusters, is defined as the number of clusters which are completely enclosed in the tested flat region. Feature #2, height indicator, represents an indicator of a height difference on the boundary of a tested flat region. Concerning height values of pixels locating in a 5 5 square window centered on each boundary pixel, the height difference is calculated by subtracting its average outside pixel heights from inside pixel heights. The height indicator is the mean value of the positive part of height difference, where the negative part is ignored since we only want to know how high the tested flat region is when it is compared to ground surface. Feature #3, area size, is simply the amount of the occupied area of a tested flat region. Feature #4, slope indicator, is obtained similarly to feature #2. Concerning those slope values of pixels which are located in a 5 5 square window centered on the each boundary pixel, the mean slope of outside the tested flat region is calculated. The positive/negative sign of slopes is assigned along the boundary by the sign of height differences obtained during the estimation of feature #2. The slope indicator is the mean va lue of signed slopes. Based on the above features, a tested flat region or a cluster is classified as a ground cluster or an object cluster by the rule-based methods showed in Figure 6-5. The feature maps of each cluster and the classification result at the site22 are shown for an example as Figure 6-6. The obtained DTM is shown in Figure 6-7a to compare with the reference DTM in Figure 6-7b created from the ground tr uth. In addition, the initial DTM and reference DTM for all of the study sites provided by ISPRS ar e shown in Figure 6-8 through Figure 6-21. Any missing objects will easily be noticed by comparing ou r initial DTM to the reference DTM since they are higher than the ground surface. By observing these figures, it is obvious that almost all objects are detected and removed in this stage. This grid-based object detection algorithm is developed here to be a pre-filter for the further processing of the st atistical bare-Earth extraction algorithm.

PAGE 78

78Statistical LiDAR Filtering Chi-Distribution Measurement For different terrain types such as a city or a forest terrain, the height differences between neighbored pixels could be varied, but the slope differences among various terrain types should be similar and in a small range. Since most objects are re moved from the method in the previous section, we are able to calculate the eight slope differences iS(8 ,..., 2 1 i) between any pixel and its 8-connected neighbored pixels (labeled by Table 6-3) from the initial terrains. Each slope difference, iS(8 ,..., 2 1 i), is assumed to be a zero-mean normal distribution in the reference DTM from the ground truth. Although the histograms of obtained slope differences from the initial terrain of the site22 (Figure 6-22) look like zero-mean normal distributions, some necessary steps still need to be done to get a more accurate terrain Under our assumption, the slope variation consisting of 8 slope differences is a 8 degree of freedom Chi distribution described by 8 1 2 i i ChiS S. (6.5) For a Chi distribution, the degree of freedom k can be expressed in terms of its mean and standard deviation by 2 2 k When the degree of freedom is equal to 8, we should get 82 2 k. A novel algorithm is proposed as Figure 6-23, where 8 k is the degree of freedom. This algorithm works as follows. First, the sum of squared mean and squared standard deviation is calculated. If it is greater than the degree of freedom the absolute maximum slope difference of the point cloud is removed, which should be an object point. To separate all object points, these steps are iteratively executed while k 2 2 Finally, the slope difference threshold TS between ground and non-ground surface is obtained by finding the abso lute maximum from remaining slope differences. Using the site22 as an example, Figure 6-24 shows its estimated degree of freedom from the assumed Chi distribution of slope differen ces is decreasing with our iterations. The TS is obtained and

PAGE 79

79 equal to 3.0 when k 2 2 where = 2.66, standard deviation = 0.99, and k = 8.06. The histogram of remaining slope differences is showed in Figure 6-25 which tends to be a Chi distribution. Adaptive Height Threshold Derivation With the obtained slope difference threshold TS, we are able to derive the adaptive height threshold which indicates the allowed growing height z z h for each pixel. Considering the 8-connected pixels give different allowed growing heights, the xif and yif are changed by adding ih to g h a f fi i xi xi6 (6.6) g h b f fi i yi yi6 (6.7) where i i iz z h and iaand ib are listed in Table 6-4. The i represents the associated position between the center pixel and its 8-connected pixels which can be found in Table 6-1. The variable ic shown in Table 6-4 is for later reference. Then, the new slope iSis changed to 2 2 2 26 6 arctan arctan g h b f g h a f f f Si i yi i i xi yi xi i (6.8) By combining the condition 0 ih if 2 2arctanyi xi i if f S S the allowed growing height ih can be solved as otherwise R h if c g S S c d d d hi i i i i i i i, 0 6 tan tan sgn2 2 2 (6.9) where TS S S yi i xi i if b f a d and ic is shown in Table 6-4. In the case where R hi there is no space for growing which leads the allowe d growing height to be 0. The allowed growing

PAGE 80

80 height threshold for each pixel is limited by its 8connected slope differences which result in 8-connected derived height limits. The minimum of these limits could underestimate ground points and restrain the growing space of ground points to cause the Type I error; while the maximum of these limits could overestimate ground points to cause the Type II error. Therefore, the adaptive height threshold of each pixel is determined by the mean func tion of its 8-connected height limits 9 5 1, max 8 1i i i i hh h T. (6.10) Accordingly, the points in the ground and object cl uster can simply be reclassified by their AGL values. If a ground point has an AGL that is higher than its corresponding threshold hT, this point is reclassified as an object point. Conversely, if an object point has an AGL that is lower than its corresponding threshold hT, this point is reclassified as a ground point. This reclassification procedure needs iteration to get the final terrain. The iterate d process ends when the total number of corrected ground and object points of the current iterati on is not reduced from the previous iteration. Grid-Based Bared-Earth Extraction Result The result from this slope-based statistical approach is shown in Table 6-5. In addition, the method is also compared to the previous approaches Cha ng#1 and Chang#2 in Table 6-6. It shows that the slope-based statistical algorithm (SSA) is better than previous results in average mean and kappa index of agreement. In addition, the standard deviation of kappa index of agreement is also less than the others which means its performance is more stab le when applied to various sites. Summary LiDAR is an important modality in terrain and land surveying for many environmental, engineering and civil applications. Recently, an unsupervised cl assification algorithm called Skewness Balancing was developed for object and ground point separation in airborne LiDAR data [76, 77]. Although the main advantages of their algorithm are threshold-freedom and independence from LiDAR data resolution, they have to build a prediction model to categorize LiDAR tiles as hilly or moderate terrains. However, not all

PAGE 81

81 LiDAR data can be categorized as either completely hilly or moderate terrain tiles. Once a tile includes both terrain types, their algorithm will face a big challenge. Our slope-based statistical algorithm is appropriate to any mixing or complicated terrain types. Most objects are removed and initial terrains are obt ained in the object detection algorithm. Slope differences are almost similar and assumed to be a zeromean normal distribution in all kinds of terrains, unlike absolute height information used by Skew ness Balancing algorithm. Based on slope difference variations, the Chi distribution measurement is used to decide the adaptive slope threshold. Accordingly, the adaptive growing height threshold of each pixel is derived by 8-connected neighbored pixels. Finally, we demonstrate the performance of this novel algorithm by testing 15 study sites from ISPRS. It shows that this algorithm is better than algorithms Cha ng#1 and Chang#2 which have outperformed all other twelve algorithms working on the same study sites shown in Chapter 5.

PAGE 82

82 Table 6-1. Associated z positions in the slope formula 7 8 9 4 5 6 1 2 3 Table 6-2. Kernel weights of the smoothing low-pass filter 1 2 1 2 4 2 1 2 1 Table 6-3. Related positions of 8-connected neighbored pixels 7 6 5 8 4 1 2 3 Table 6-4. Variable list for the 8-connected pixels i ia ib 2 2i i ib a c 1 -1 -1 2 2 0 -1 1 3 1 -1 2 4 -1 0 1 6 1 0 1 7 -1 1 2 8 0 1 1 9 1 1 2 Table 6-5. Total errors and kappa index for 15 study sites by the slope-based statistical algorithm (SSA). Site TP TN FNFPErrorKappa Samp11 19306 2480 16211460310.8%78.1% Samp12 25833 858 637247912.9%94.3% Samp21 9935 150 7827971.8%94.9% Samp22 21984 520 58096223.4%92.2% Samp23 12607 616 558113144.7%90.6% Samp24 5277 157 23518235.2%86.7% Samp31 15493 63 233130731.0%97.9% Samp41 5451 151 19454353.1%93.9% Samp42 12061 382 210298171.4%96.6% Samp51 13901 49 57433213.5%89.2% Samp52 19894 202 41119492.7%84.9% Samp53 32498 409 4059812.4%69.4% Samp54 3832 150 22044054.3%91.4% Samp61 33784 54 3388671.1%81.0% Samp71 13797 76 27514952.2%88.2%

PAGE 83

83 Table 6-6. Performance comparison among different methods Method Error mean Kappa mean Kappa Standard Deviation Chang#1 4.6% 84.5%12.1% Chang#2 3.8% 87.3%10.2% SSA 3.4% 88.6%7.7%

PAGE 84

84 Figure 6-1. Standard deviation and confid ence interval of a normal distribution. x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 5 10 15 20 25 30 35 Figure 6-2. Outlier points found by a normal distribu tion with 95% confidence interval at site22: 20 (black) points and 49 (white) points are lower and higher than the interval. x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 10 20 30 40 50 60 70 80 A x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 0 1 2 3 4 5 6 7 8 B Figure 6-3. Edge detection at the site22: A) Smoothed slope map and B) smoothed tophat filtering at the site22

PAGE 85

85 x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 Figure 6-4. Obtained flat (black) regions at the site22. Figure 6-5. Rule-based methods for flat region classification. #inside clusters<10 Ground cluster Rule#1: Area size> 350 & slope indicator<0 Rule#2: Area size>100 & slope indicator<-20 Rule#3: Area size>50 & slope indicator<-30 Pass any rule? Height indicator<2.5m Object cluster Each flat region Yes No No Yes Yes No

PAGE 86

86 x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 0 10 20 30 40 50 A x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 0 1 2 3 4 5 6 7 8 9 10 B x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 -10 -8 -6 -4 -2 0 2 4 6 C x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 -80 -60 -40 -20 0 20 40 60 D x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180 200 E x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 0 0.2 0.4 0.6 0. 8 1 1.2 1.4 1.6 1. 8 2 F Figure 6-6. The associated feature maps and classification result of the site22: A) cluster labeling map B) cluster-wise inside cluster number map C) cluster-wise positive average height map D) cluster-wise signed average slope map E) cluster-wise area size map F) classification result.

PAGE 87

87 x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 4 6 8 10 12 14 16 18 A x (meter)y (meter) 20 40 60 80 100 120 140 160 180 20 40 60 80 100 120 140 160 180 4 6 8 10 12 14 16 18 B Figure 6-7. DTM comparison at the site22: A) initial DTM and B) reference DTM x (meter)y (meter) 20 40 60 80 100 120 50 100 150 200 250 300 0 10 20 30 40 50 60 70 80 90 100 A x (meter)y (meter) 20 40 60 80 100 120 50 100 150 200 250 300 0 10 20 30 40 50 60 70 80 90 100 B Figure 6-8. DTM comparison at the site11: A) initial DTM and B) reference DTM x (meter)y (meter) 50 100 150 200 50 100 150 200 250 74 76 78 80 82 84 86 88 90 92 94 A x (meter)y (meter) 50 100 150 200 50 100 150 200 250 74 76 78 80 82 84 86 88 90 92 94 B Figure 6-9. DTM comparison at the site12: A) initial DTM and B) reference DTM

PAGE 88

88 x (meter)y (meter) 20 40 60 80 100 120 10 20 30 40 50 60 70 80 90 100 110 0 0.5 1 1.5 2 2.5 3 3.5 A x (meter)y (meter) 20 40 60 80 100 120 10 20 30 40 50 60 70 80 90 100 110 0 0.5 1 1.5 2 2.5 3 3.5 B Figure 6-10. DTM comparison at the site21: A) initial DTM and B) reference DTM x (meter)y (meter) 20 40 60 80 100 120 140 20 40 60 80 100 120 140 160 180 200 25 30 35 40 45 50 A x (meter)y (meter) 20 40 60 80 100 120 140 20 40 60 80 100 120 140 160 180 200 25 30 35 40 45 50 B Figure 6-11. DTM comparison at the site23: A) initial DTM and B) reference DTM x (meter)y (meter) 20 40 60 80 100 120 10 20 30 40 50 60 70 0 5 10 15 20 A x (meter)y (meter) 20 40 60 80 100 120 10 20 30 40 50 60 70 0 5 10 15 20 B Figure 6-12. DTM comparison at the site24: A) initial DTM and B) reference DTM

PAGE 89

89 x (meter)y (meter) 20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160 81.5 82 82.5 83 83.5 84 84.5 85 85.5 86 A x (meter)y (meter) 20 40 60 80 100 120 140 160 20 40 60 80 100 120 140 160 81.5 82 82.5 83 83.5 84 84.5 85 85.5 86 B Figure 6-13. DTM comparison at the site31: A) initial DTM and B) reference DTM x (meter)y (meter) 20 40 60 80 100 120 140 160 20 40 60 80 100 34 36 38 40 42 44 A x (meter)y (meter) 20 40 60 80 100 120 140 160 20 40 60 80 100 34 36 38 40 42 44 B Figure 6-14. DTM comparison at the site41: A) initial DTM and B) reference DTM x (meter)y (meter) 50 100 150 200 20 40 60 80 100 120 140 160 180 200 0 1 2 3 4 5 6 7 8 9 A x (meter)y (meter) 50 100 150 200 20 40 60 80 100 120 140 160 180 200 0 1 2 3 4 5 6 7 8 9 B Figure 6-15. DTM comparison at the site42: A) initial DTM and B) reference DTM

PAGE 90

90 x (meter)y (meter) 50 100 150 200 50 100 150 200 250 300 350 400 -5 0 5 10 15 20 25 30 35 40 A x (meter)y (meter) 50 100 150 200 50 100 150 200 250 300 350 400 -5 0 5 10 15 20 25 30 35 40 B Figure 6-16. DTM comparison at the site51: A) initial DTM and B) reference DTM x (meter)y (meter) 50 100 150 200 250 300 350 400 450 50 100 150 200 250 300 0 20 40 60 80 A x (meter)y (meter) 50 100 150 200 250 300 350 400 450 50 100 150 200 250 300 0 20 40 60 80 B Figure 6-17. DTM comparison at the site52: A) initial DTM and B) reference DTM x (meter)y (meter) 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 450 0 10 20 30 40 50 60 70 A x (meter)y (meter) 50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400 450 0 10 20 30 40 50 60 70 B Figure 6-18. DTM comparison at the site53: A) initial DTM and B) reference DTM

PAGE 91

91 x (meter)y (meter) 50 100 150 50 100 150 200 250 25 30 35 40 45 50 A x (meter)y (meter) 50 100 150 50 100 150 200 250 25 30 35 40 45 50 B Figure 6-19. DTM comparison at the site54: A) initial DTM and B) reference DTM x (meter)y (meter) 100 200 300 400 500 50 100 150 200 250 300 350 400 0 5 10 15 20 25 30 35 A x (meter)y (meter) 100 200 300 400 500 50 100 150 200 250 300 350 400 0 5 10 15 20 25 30 35 B Figure 6-20. DTM comparison at the site61: A) initial DTM and B) reference DTM x (meter)y (meter) 50 100 150 200 250 300 350 50 100 150 200 2 4 6 8 10 12 1 4 A x (meter)y (meter) 50 100 150 200 250 300 350 50 100 150 200 2 4 6 8 10 12 14 B Figure 6-21. DTM comparison at the site71: A) initial DTM and B) reference DTM

PAGE 92

92 -20 -10 0 10 20 0 1000 2000 3000 Histogram of S1: =0.15, = 3.05 -20 -10 0 10 20 0 1000 2000 3000 Histogram of S2: =0.10, = 2.22 -20 -10 0 10 20 0 1000 2000 3000 Histogram of S3: =0.41, = 3.90 -20 -10 0 10 20 0 1000 2000 3000 4000 Histogram of S4: =0.22, = 2.91 -20 -10 0 10 20 0 1000 2000 3000 Histogram of S5: =0.26, = 3.06 -20 -10 0 10 20 0 1000 2000 3000 4000 Histogram of S6: =0.09, = 2.26 -20 -10 0 10 20 0 1000 2000 3000 Histogram of S7: =0.26, = 3.70 -20 -10 0 10 20 0 1000 2000 3000 4000 Histogram of S8: =0.10, = 2.87 Figure 6-22. Histograms of slope differences from the initial terrain at the site22

PAGE 93

93Load LiDAR point cloud while k 2 2 do Remove absolute maximum slope difference end while Get slope difference threshold, TS, between ground and non-ground surface by finding absolute maximum from remaining slope differences. Figure 6-23. Adaptive slope difference threshold algorithm based on Chi distribution 0 100 200 300 400 500 600 700 800 100 101 102 Iterationsk, degrees of freedomChi distribution of slope difference vs. iterations k= 2+ 2 Figure 6-24. The change curve of freedom from slope differences with iterations at the site22 0 1 2 3 4 5 6 7 0 50 100 150 200 250 300 350 400 450 Figure 6-25. Histogram of remaining slope differences which tends to be a Chi distribution at the site22

PAGE 94

94CHAPTER 7 OCCLUDED VEHICLE DETECTION DESIGN Vehicle detection has been utilized in the In telligent Transportation System (ITS), Automatic Vehicle Guidance (AVG), and traffic flow estimation, but has not been exploited in the forested terrain. LiDAR employs an active optical modality or laser ranging that provides primarily geometric information to detect natural surface features and other hard targets that may be spectrally inseparable in multi-spectral passive optical imagery. In addition, airborne LiDAR can provide data in large spatial extents with varying temporal resolution and it can be deployed more or less anywhere and at any time which restricts the use of passive optical imagery in cluding in smoke, haze, f og, and at night. Thus, occluded vehicle detection from airborne LiDAR data in forested terrain can be applied to many fields, more specifically: 1) military surveillance – searchi ng enemy vehicles in a battle area with forest, 2) homeland security – border crossing monitoring for vehi cles in forest area, 3) global warming – vehicle hunting for illegal deforestation which is a hidden cause of global warming, 4) disaster rescue – finding vehicles stuck by disrupted roads in forest during natural disaster, 5) emergency road service – locating vehicles involved with general car accidents in forest, and 6) criminal searching – uncovering forest canopy to search suspicious vehicles hiding in mountains. The occluded objects underneath trees can be rev ealed though the deve loped canopy removal algorithm in the beginning. The obtained uncovered Li DAR points are clustered into individual objects by the proposed DTM extraction algorithm and th e associated morphological image processing, including consideration of horizontal and vertical orientations. The clustered LiDAR points of each object will be exploited by many theories such as Spin image, non-parameteric Parzen-window estimation, Bayesian decision, and relative entropy, etc. A probabilistic modeling is built up to detect those occluded vehicles in forested terrain from ai rborne LiDAR point clouds. Finally, we verify our results by examining the Receiver Op erating Characteristic (ROC) curves.

PAGE 95

95Object Clustering Cluster analysis, or clustering, is the assignment of a set of observations into subsets (called clusters) so that observations in the same cluster are similar in some sense. Clustering is a method of unsupervised learning, and a common technique for statistical data analysis used in many fields, including machine learning, data mining, pattern recognition, image anal ysis and bioinformatics [78]. In the application of image segmentation, clustering can be used to divide a digital image into distinct regions for border detection or object recognition. Object recognition in computer vision is the task of finding a given object in an image or video sequence. Humans recognize a multitude of objects in im ages with little effort, despite the fact that the image of the objects may vary somewhat in different view points, in many different sizes/scale or even when they are translated or rotated. Objects can ev en be recognized when they are partially obstructed from view. This task is still a challenge fo r computer vision systems in general [79]. A good approach for object recognition is the scale-invariant feature transform (or SIFT) which is an algorithm in computer vision to detect and describe local features in images. SIFT key points of objects are first extracted from a set of reference imag es and stored in a database. An object is recognized in a new image by individually comparing each feat ure from the new image to this database and finding candidate matching features based on Euclidean distance of their feature vectors. From the full set of matches, subsets of key points that agree on the object and its location, scale, and orientation in the new image are identified to filter out good matches. The determination of consistent clusters is performed rapidly by using an efficient hash table implementa tion of the generalized Hough transform. Each cluster of 3 or more features that agree on an object and its pose is then subject to further detailed model verification and subsequently outliers are discarded. Finally the probability that a particular set of features indicates the presence of an object is computed, given the accuracy of fit and number of probable false matches. Object matches that pass a ll these tests can be identified as correct with high confidence [80]. In addition, k-means clustering [81] is a met hod of cluster analysis which aims to partition n

PAGE 96

96 observations into k clusters in which each observation belongs to the cluster with the nearest mean. The k-means clustering algorithm is commonly used in computer vision as a form of image segmentation. The results of the segmentation are used to aid border detection and object recognition.The two key features of k-means which make it efficient are often regarded as its biggest drawbacks: 1) the number of clusters k is an input parameter and an inappropriate choice of k may yield poor results; 2) Euclidean distance is used as a metric and variance is used as a measure of cluster scatter. However, our proposed object clustering method does not have the need to create a database from the training data in the SIFT or to choose the number of clusters k and use the Euclidean distance as a metric in the k-means clustering. It is important to obtain the whole shape of each object by clustering analysis for identification, but it is a challenge to do when vehicles are mixed with other objects and irregular canopy occlusion will more or less redu ce reflected points from underneath objects. Thus, it will make different objects which are close to one another become hardly seg mented since the gaps between objects could be occluded such that the number of points reflected from the gaps become impacted. Therefore, we developed a feasible way for the object segmentation (Figure 7-1), which includes horizontal-based and vertical-based morphological filtering, following our canopy removal algorithm and ground point filtering algorithm. Horizontal Based Morphological Filtering First, the DTM is a regular grid image which can be generated by resampling irregular ground points extracted from our ground point filtering, where th e grid resolution is 5 pixels by 5 pixels per square meter which was approximately equal to our average LiDAR density. By referring to the DTM, the individual subtractions with corresponding elevati on of DTM from the original height of points are used to obtain the above ground levels (AGL) of all non-canopy points, which were remaining points after removing canopy points above vehicles. Then the DSM is generated by resampling irregular non-canopy points with their AGL values in the same grid resolution as the DTM. Finally, two kinds of segmentations are followed to cluster individual object points: horizontal-based and vertical-based morphological filtering. Some summary of morphologi cal filtering operations is shown in Table 7-1.

PAGE 97

97 The horizontal-based segmentation is used to separate object points which are not close enough in the horizontal direction or x-y coordinates to fo rm an individual object. The opening operator is performed first by a disk-shaped structuring element (SE) with a radius of two pixels on the DSM. As such, those objects are partially conn ected and their lower parts can be segmented away. The disk type of structure element is chosen as it is orientation invari ant and the orientations of objects could be in any direction. In the seriously occluded situation, the partial c onnection between two objects could be larger than the SE of the opening operator to make the above seg mentation fail. However, the bot-hat operator is useful for enhancing detail in the presence of local bottoms [61]. Thus, the bot-hat operator using a disk-shaped SE with a radius of ten pixels is applied to find a channe l-like structure where the center is lower than its neighbor. Using both opening and bot-hat morphological f ilters we are able to segment two close objects with short or long partial connections. Therefore, this spacing-based segmentation is very helpful to separate those vehicles which are close to one another in the parking lot. Vertical Based Morphological Filtering The vertical-based segmentation is applied to sep arate those points which are close in vertical direction, or z coordinate, but belong to differen t objects. In the process of removing tree canopy above vehicles, some vegetation points lower than the maxi mum height of vehicles still remained since they could be parts of potential vehicles and need to be fu rther investigated. Besides, some vehicles could be close to buildings and the thin gaps between them could not reflect any points due to irregular tree canopies. Both cases need segmentation in the vertical direction. We proposed a simple but effective method consisting of the horizontal-based segmentation described above with two elevation limitations. The first limit of elevation is to filter the DSM by examining and removing those pixels higher than 2 mete rs, which is referred to as the average heights of sedans, SUVs, and pickup trucks. This modified DSM is then applied to the horizontal-based segmentation with the same opening and bot-hat morphological filters. In this way, those noise points located above vehicles can easily be removed to help with further vehicle identification.

PAGE 98

98 Buses, 18-wheel trucks and military equipment (e.g ., tanks) are higher than 2 meters but usually lower than 4 meters. Therefore, if these large vehicl es exist in the dataset, then another DSM has to be obtained by removing those pixels higher than the second limit of elevation, 4 meters. Similarly, the spacing-based segmentation above is followed agai n to achieve the elevation-based segmentation. Accordingly, those points belonging to one object can be clustered by the object segmentation from both spacing-based and elevation-based morphological filtering methods. Object Clustering Result Figure 7-2, Figure 7-3, and Figure 7-4 show the r esults of object segmentation in Hogtown parking site, Hogtown forest site, and the residential site, respectively. Comparing these to Figure 4-11b, Figure 4-12b, and Figure 4-13b, those segmented objects achieve a promised accomplishment since nearly all points belonging to vehicles are clustered indi vidually and segmented with other non-vehicles. When considering the corresponding LiDAR density in these three study sites, it was calculated that the average LiDAR densities after removing canopy points are 15.77, 9.73, and 18.98 points per square meter and the standard deviation is 8.68, 6.45, and 8.41 for Hogtown parking site, Hogtown forest site, and the residential site, respectively. The en tire space distribution for LiDAR density in these three study sites can be seen in Figure 4-14a, Figure 4-14b, and Figure 4-14c. However, it is noted that even if the LiDAR points are from the laser scanning swath of only one flight pass, in which the point density could be only one sixth or one seventh of the cu rrent density, then this object segmentation method combining spacing-based and elevation-based morphologi cal filtering still can work well as is shown in a later section. Object Classification After clustering points as segmented objects, the whol e shape of individual objects can be formed and their special features can be extracted and analyzed for this object classification: either a vehicle or a non-vehicle class. If vehicles were in an open area which was not occluded by trees, the vehicle/non-vehicle classification could be much easier such as in typical research for estimating traffic flow. For occluded vehicles underneath trees, two problems emerge for vehicle/non-vehicle classification:

PAGE 99

99 the intra-class problem and inter-class problem. The fo rmer problem is that even if the vehicle size is variable due to not only vegetation occlusion but al so different types, manufacturing companies, and models, all un-occluded parts of different vehicles ha ve to be recognized as the vehicle class. The latter problem is how to distinguish those whole or pa rtial vehicles with variable size from variable non-vehicle objects. We propose the following appr oach to solve both intra-class and inter-class problems simultaneously. The block diagram of our object classification is shown in Figure 7-5. Principal Component Analysis Principal component analysis (PCA) [82] involv es a mathematical procedure that transforms a number of possibly correlated variables into a smalle r number of uncorrelated va riables called principal components. The first principal component accounts for as much of the variability in the data as possible, and each succeeding component accounts for as much of the remaining variability as possible. PCA is the simplest of the true eigenvector-based multivariate analyses. Often, its operation can be thought of as revealing the internal structure of the data in a wa y which best explains the variance in the data. If a multivariate dataset is visualized as a set of coordi nates in a high-dimensiona l data space (1 axis per variable), PCA supplies the user with a lower-dime nsional picture, a “shadow” of this object when viewed from its (in some sense) most informative viewpoint. Component analysis is an unsupervised approach to finding the “right” features from the data. In PCA [83], we usually represent the d-dimensional data in a lower-di mensional space. This will reduce the degrees of freedom, reducing the space and time co mplexities. The common goal is to represent data in a space that best describes the va riation in a sum-squared error sense. One example of using PCA is to get the best fitting plane from irregular 3-dimensi onal points. Figure 7-6 shows that the first two principal components are obtained which are orthogon al each other and can define vectors that form a basis for the plane. The third PC is also orthogonal to the first two, and it defines the normal vector of the plane. We apply PCA to find the length and width of each object. First, each object consists of clustered non-canopy points with their AGL heights. All vehi cles can be considered as rectangle shapes when

PAGE 100

100 projecting to x-y coordinates. The length and widt h of vehicles can be a good pre-filter to remove non-vehicle objects. Since the orientation of every object is arbitrary, PCA [83] is a good tool to get its orientation and find both length and width of any object. Let v be the N 2 matrix consisted of N points with x-y coordinate pairs. The 2-D mean vector and 2 2 covariance matrix are computed for the data set v Next, the eigenvectors and eigenvalues are computed and sorted according to decreasing eigenvalue. Then, the 2 2 matrix A is formed whose columns consist of the 2 eigenvectors. The representation of data by principal components consists of projecting the data onto 2-D according to [83] v A vt, (7.1) where v consists of two columns x and y Therefore, the length, L and width, W of an object can be obtained by x x min maxL (7.2) y y min max W (7.3) Taking account of the original vehicle size and acceptable error margin, the length and width of vehicle candidates should be less than 8 meter and 4 me ter, respectively. Thus, if the length or width of one object is greater than the corresponding threshold, it will be classified as a non-vehicle object, while those vehicle candidates which satisfied the length an d width criterions continue to be examined and their features will be extracted from the Spin image. Spin Image The spin image [84] is a method of surface matc hing which is the process that compares surfaces and decides whether they are similar. Surface matchi ng can be used for object recognition by analyzing and comparing the features extracted from object surfaces. Surface matching is difficult because the coordinate system in which to compare two su rfaces is undefined, characteristics of sensed data, including clutter, occlusion and sen sor noise. Surface matching is furt her complicated in sparse LiDAR point density situations. The spin image describes a data level representati on of surfaces used for surface matching. In its

PAGE 101

101 representation, surface shape is described by a collect ion of oriented points (3-D points with a surface normal). Using a single point basis constructed from an oriented point, the position of other points on the surface can be described by two parameters. The accumulation of these parameters for many points on the surface results in an image at each oriented poi nt. These images, localized descriptions of the global shape of the surface, are invariant to rigid tran sformations. Through corre lation of images, point correspondences between two surfaces can be estab lished. When two surfaces have many point correspondences, they match. Taken together, the oriented points and associated images make up the surface representation. Because the image generation pr ocess can be visualized as a sheet spinning about the normal of a point, the images in this kind of representation are called spin-images. The two coordinates of the basis in spin-images are the perpendicular distance to the line L and the signed perpendicular distance to the plane P An oriented point basis is a cylindrical coordinate system that is missing the polar angle coordinate becau se this coordinate cannot be determined using just surface position and normal. Usi ng an oriented point basis O a spin-map SO is defined as the function in [84] that projects 3-D points x to the 2-D coordinates of a particular basis ( p n ) corresponding to oriented point O 2 3: R R SO (7.4) p x n p x n p x x SO, ,2 2 (7.5) Then for each vertex x on the surface of the object, the spin -map coordinates with respect to O are computed and bilinearly interpolated to form a digita lized 2-D array. The pixel value which each point x is spin-mapped would be to increment by one. Once a ll of the points on the surface of the object have been accumulated, a 2-D array representation of the spin-image is generated. In our vehicle detection, the oriented plane is chosen as the local ground surface of the mapped object, where its normal vector n is orthogonal to this oriented plane, while the oriented point p is set as the center point of the object on the ground level. Then, the new 2-D coordinates (,) of the Spin map can be obtained from 3-D points of the object by Equation (7.5). In order to count points of the Spin map

PAGE 102

102 inside each bin of the Spin image, the contribution of each point is bilinearly interpolated to the four surrounding bins in the 2-D array making the array less sensitive to the position of the point, where the bin size is set as 0.2 meter, approximately equal to LiDAR vertical accuracy. Once all resampling points of the object have been accumulated to their corr esponding bins, a 2-D array representation of the Spin image is generated. Taking some examples of collected vehicle data with variable occluded degree, their 3-D shapes from irregularly reflected points are gene rated by the Delaunay triangulation to see how those vehicles were seriously impacted by the occlusion, and their corresponding 2-D Spin images are obtained and shown in Figure 7-7 through Figure 7-16, sorted by the actual collected number of LiDAR points from individual vehicles. Although irregularly occluded vehicles are variable in length, width, height, and shape, they could be transformed to the Spin images, SI(,), to analyze and extract some important features. By our observation and analysis on those Spin im ages of vehicles, two special features, and are extracted according to n SI 2 1 max arg (7.6) nn11 (7.7) 111 2SInm (7.8) where nmSI11, (7.9) 0 : max SI n, (7.10) and 0 : max SI m. (7.11) The individual locations of are marked by black circles for illustration in the Spin images of Figure 7-7 to Figure 7-16, where all locations of are connected by piecewise lines. By using the spin

PAGE 103

103 image, points are accumulated and counted if they have the same horizontal distance, and the same vertical distance, away from the oriented point. The spin im age benefits the vehicle detection in the following two ways: 1) we can take advantage of the symmetry of a vehicle where the shape of the right side is the same as the left side in the viewpoint of the front or back; 2) the spin image is rotation invariant which avoids the need to detect the orie ntation of diverse and occluded vehicles. Even though PCA can be helpful to determine the orientation of vehicles, it becomes more difficult when facing the situation with sparser point density. Therefore, we can feasibly apply the spin image to LiDAR data on the vehicle detection in a sparse point density and an arbitrary vehicle orientation. Using the extracted features and of vehicle candidates including vehicles and non-vehicles, two corresponding bivariate non-parame tric PDFs (Probability Density Function) can be estimated by the Parzen windowing method with a Gaussian kernel fu nction [49] from the training data set. The optimal window size selection can be determined by Silverman’s formula [85], 4 11 2 4 d x optN d (7.12) where d is the data dimensionality, N is the sample size of the data, and i X xiid1, whereiiXare the diagonal elements of the sample covari ance matrix. Accordingly, the bivariate PDF of vehicles and non-vehicles constructed from two features and is obtained and shown in Figure 7-17 and Figure 7-18, respectively. Bayesian Decision Bayesian decision theory is a fundamental sta tistical approach to the problem of pattern classification by using probability and the costs associat ed to decisions. The Bayesi an decision rule [83] is to decide 1 if 1 11 21 2 22 12 2 1 P P x p x p (7.13)

PAGE 104

104 where j i ij is the loss incurred for deciding i when the true state of nature is j jx p is a function of j (i.e., the likelihood function) and jx p is a priori probability of j Thus the Bayesian decision rule can be interpreted as calling for deciding 1 if the likelihood ratio, 2 1 x p x p, exceeds a threshold value represented by the term in the right hand side of Equation (7.13) which is independent of the observation x In our study, 1 and2 is the class for vehicle and non-vehicle, respectively. Given a test object with two special features and the probability in the vehicle/non-vehicle category can be found from the obtained bivariate PDF of vehicles/non-vehicl es. If the probability in the vehicle category is higher than in the non-vehicle category or the ratio of th e former over the latter is greater than 1, then the test object is classified as a vehicle; otherwise, it is classified as a non-vehicle. Relative Entropy The relative entropy or Kullback-Leibler distan ce is a measure tool to calculate the distance between two distributions. It is strongly relate d to information divergence and information for discrimination which discrete version is defined as [83] x KLx p x q x q x q x p D ln (7.14) where p ( x ) and q ( x ) are two discrete distributions over the same variable x Its continuous version [83] is dx x p x q x q x q x p DKL ln (7.15) It is noted that the 0 q p DKL and 0 q p DKL if and only if q p. In our study, we apply this relative entropy to the Ba yesian decision. First, the right hand side of the Bayesian decision equation can be re garded as a Bayesian threshold ( BT ). BT P P x p x p 1 11 21 2 22 12 2 1 (7.16)

PAGE 105

105 Since we want to get optimal decision for vehicles, the PDFs between referenced vehicles and classified vehicles should be as similar as possible. The relativ e entropy can be used to measure the distance between the distributions of referenced vehi cles and classified vehicles. With varying the BT in a smart way, such as where one tests the large scale first to find a good threshold and obtain small range for the small scale trial, the various relative entropy measu rements can be obtained for vehicle detection. Of course, the smaller the relative entropy is, the better the BT is. However, the Bayesian decision is based on quan tifying the tradeoffs between various classification decisions using probability. Similarly, our approach cannot ignore the relative entropy for non-vehicle detection in order to achieve the optimal decisi on for considering both vehicles and non-vehicles. Therefore, the best trade off, or the optimal BT value, is achieved by considering two relative entropies between vehicle and non-vehicle detection. In our case, the relative entropy between two di stributions of vehicles and non-vehicles based on extracted features and is fixed. But, the distribution of detected vehicles/non-vehicles can be changed with the adjustable variable BT such that the relative entrop y between two distributions of reference vehicles and detected vehicles is varied, and so is that between reference non-vehicles and detected non-vehicles. where these two situations ar e shown as the curves for vehicles and non-vehicles, respectively, in Figure 7-19. Therefore, the relativ e entropy can help to examine and evaluate which BT can make our detected distribution of vehicle/non-vehi cle be as close as possible to the true distribution of vehicle/non-vehicle. After measuring the KLD information divergences between reference and detection in both vehicles and non-vehicles, it is observed in Figure 7-19 that two curves,KLD functions of the BT for vehicles and non-vehicles approach each other initia lly and diverge gradually in the sense of ascending BT The detectability for non-vehicles gradually becom es worse but it turns better for vehicle detection and then turns worse. The optimal BT here is one that achieves the best overall detection for both vehicles and non-vehicles. If the BT is solved by finding the lowest average of summing two KLDcurves,

PAGE 106

106 it cannot be guaranteed as the best solution because th e averaging is equal to giving the same weighting coefficients while the prior probabilities of vehi cles and non-vehicles are different and probably unknown. Instead of the summing method, the subtracting method is proposed. From Figure 7-19, it is observed that the detectability for non-vehicles is al ways stronger than the detectability for vehicles. Thus, the best BT should be able to suppress the detectability for non-vehicles and boost the detectability for vehicles simultaneously, which inspires the following equation. BT D BT D BTKL KL BT optNonVehicle Vehicle ,min arg (7.17) Actually, the above difference of KLD values between the vehicle and non-vehicle class is the indicator of detection bias. The larger difference means the higher preference of discriminability to non-vehicles. For example, a BT smaller than 0.4 makes the detection of non-vehicles pretty good since its correspondingKLD is almost 0, while the detection of vehicles becomes very poor which are observed by those large corresponding KLD values. Equation (7.17) can also be interpreted to give better detection for vehicles and balance the detecti on bias between vehicles and non-vehicles. Therefore, the optimal BT of 0.7 can be found from Equation (7.17). Th e bivariate PDF of detected vehicles and non-vehicles with setting BT as 0.7 is shown in Figure 7-20 and Figure 7-21, respectively. Occluded Vehicle Detection Result and Evaluation By using the Bayesian decision rule in Equation (7.16) and the optimal BT in Equation (7.17) determined by the associated information divergence, each segmented object with its extracted features and can be classified as either a vehicle or a non-v ehicle based on generated bivariate distributions ) (1 pand ) (2 p from the reference vehicle class in Figure 7-17 and reference non-vehicle class in Figure 7-18, respectively. The reference vehicle/non-vehicle bivariate distributions are trained by 1658 objects at Hogtown parking site and Hogtown forest site from individual and overlapped laser scans. The test site, Hogtown resi dential site, includes 416 objects from individual and overlapped laser scans which are never used for training and have no cont ribution to the true

PAGE 107

107 vehicle/non-vehicle bivariate distributions. However, a ll of the 2074 objects are tested and classified by the Bayesian decision rule with our optimal BT Our detected vehicle/non-vehicle objects are compared to true vehicle/non-vehicle objects by the cross matrix in Table 7-2, where the False Negative (FN) case, or Type I error, is for falsely rejecting vehicles, and the False Positive (FP) case, or Type II error, is for incorrectly accepting non-vehicles. The remaining True Positive (TP) and True Negative (TN) cases are for the correct classifications of vehicles and non-ve hicles, respectively. The detection accuracy [4] is obtained by Equation (7.18) composed of the TP, FN, FP, and TN. % 100 Accuracy TN FP FN TP TN TP (7.18) For Hogtown parking site, the result of vehicle de tection for overlapped laser scans is shown in Figure 7-22a, where the white lines and black lines represent detected vehicles and non-vehicles, respectively. Similarly, the result of vehicle detection for a single scan (scan #6) is shown in Figure 7-22b. Comparing to Figure 4-11b, the TP, FN, FP, and TN can be counted and the accuracy can be obtained by Equation (7.18). Accordingly, the overlapped a nd single-scanned accuracy are 95.4% and 85.5%, respectively. It is noted that the average under canopy LiDAR density of a single scan for this site is 2.25 points per square meter with a standard deviation of 1.24 points per square meter. Even though those vehicles are occluded by irregular tree structures and reflected by different laser scans, our vehicle detection accuracy for single scans still range from 83.8% to 88.2%. Besides, the overall accuracy by testing 947 objects from overlapped and single scans can reach 87.9% (Table 7-3). For Hogtown forest site, the result of vehicle de tection for its overlapped laser scans and single scan (scan #6) are shown in Figure 7-23. Similarly, th is result can be compared to Figure 4-12b and TP, FN, FP, and TN can be counted to obtain the det ection accuracy. The accuracies are 80.5% and 83.2% for overlapped scans and the single scan, respectively (Tab le 7-4). The average under canopy LiDAR density of single scan for this site is 1.62 points per square meter with a standard devi ation of 1.07 points per square meter. Due to the seriously occluded scen ario in this forest site, wher e its occluded rate of 80.59% is

PAGE 108

108 much higher than the 63.51% in the parking site, both available LiDAR density and detected accuracy for underneath objects are affected and reduced. In add ition, since there is only one tarpaulin settled as a mimic vehicle source in this site, the obtained in formation for detecting vehicles under this heavy occlusion should be insufficient no matter how many laser scans were rendered. Therefore, if more vehicle sources were set up and collected under this fo rest, the bivariate distribution of true vehicles can be improved and so can the vehicle detectability. Othe rwise, even if all single scan data are overlapped for detecting, its performance could be worse than that from only single scan. It is because those detailed vehicle-like objects could be incorrectly classified as vehicles if those counterparts of detailed true vehicles are missing in the statistical based decision. However, our overall accuracy by testing 711 objects from overlapped and single scans can still get 81. 0% (Table 7-4) due to the large amount of training data from 13 LiDAR scans. For Hogtown residential site, the vehicle det ection results for overlapped scans and single scan (scan #3) are shown in Figure 7-24, where the accur acies are 94.7% and 83.3%, respectively (Table 7-5). Obviously, this performance is much better than that in the Hogtown forest site. It is the degree of occlusion that makes the major difference of det ection, where the average under canopy LiDAR density of single scan for this site is 3.16 points per square meter with a standard devi ation of 1.40 points per square meter. Although the average under-canopy LiDA R density is a little higher than that in Hogtown parking site, the performance does not become better. A major reason is that all the objects in this site have never been used for learning in the bivariat e probability density functio ns of vehicle class and non-vehicle class. It also explains why the overall accur acy for this site drops to 80.8% (Table 7-5). In addition, its standard deviation of under-canopy LiDAR density is larger than the other sites which can explain the larger dynamic range of the detection accuracy from 71.2% to 94.7% (Table 7-5). On the other hand, the average detection accuracy is examined based on the number of collected LiDAR points from each object. In this way, the quality of detection performance can be observed by the quantity of LiDAR points. First, all objects are record ed with their numbers of collected LiDAR points. Instead of mixing all objects in one site to count TP FN, FP, and TN values, they are separated into

PAGE 109

109 different groups to count those values, where each group consists of only one specific number of collected LiDAR points. Figure 7-25 shows that th e relationship between the numbers of collected LiDAR points from each object and the vehicle detection performance including TP, FN, FP, and TN in the case of a Bayesian threshold equal to 0.7. It can be observed that the TP and TN increase with increasing abscissa while FN and FP decrease with in creasing abscissa, which demonstrates that a higher number of collected LiDAR points from a test object will give the better system detectability. Accordingly, another relationship between the number of collected LiDAR points from each object and the vehicle detection accuracy, hit ratio, and false alar m in the case of Bayesian threshold equal to 0.7 can be obtained and is showed in Figure 7-26, where the hit ratio and false alarm formula can be derived from [49] as the following equations. % 100 Ratio Hit FN TP TP (7.19) % 100 False TN FP FP Alarm (7.20) Obviously, the accuracy curve in Figure 7-26 is a monotonic increasing function since the hit ratio and false alarm generally increase and decrease with in creasing abscissa, respectively. It is noted that the detection accuracy for different collected LiDAR points per object is always beyond 80%, even though the collected point number of a test object is belo w 5 points (Figure 7-26). It is because not only the points reflected from a testing object but also its neighboring ground points are imposed to extract associated features. Finally, the optimal Bayesian threshold is exam ined by the receiver operating characteristic (ROC) curves [49], where the system performance is presented by the false alarm and hit ratio simultaneously. The ROC (Receiver Operating Characteristics) curves of our detection performance are shown for 5, 10, 15, and 20 sampling points on one object and 0 sampling points for reference in Figure 7-27. For each curve excluding the reference curve, th e Bayesian thresholds 0.5, 0.6, 0.7, 0.8, 0.9 and 1.5 are marked in ascending order. It is observed that the Bayesian threshold 0.7, or the third mark counted from the bottom

PAGE 110

110 up of each curve consisting of 7 piece-wise lines, is th e turn point, where the slopes of lines before this point is greater than 45 while the slopes of lines after this point is smaller than 45 This phenomenon explains that the optimal Bayesian threshold is 0.7. Although this ROC curve can examine the system performance to get the optimal Bayesian threshold, the procedure to obtain those TP, FN, FP, and TN values for hit ratio and false alarm is time consuming si nce each classified object has to be verified as the TP, FN, FP, or TN case. However, using the proposed method of relative entropy is an efficient and effective way to analyze and get the optimal Bayesian threshold since the vehicle/non-vehicle distribution can be created directly from the classifi ed objects without having the need to check the recognition result of each classified object. Summary We demonstrate that the state-of-the-art airborne LiDAR system can provide valuable data which can effectively support the occluded vehicle detecti on in forest terrain. The proposed system is a probabilistic model built through the canopy re moval algorithm, DTM extraction algorithm, morphological image processing, PCA, Spin image, Parzen-window estimation, Bayesian decision, and relative entropy, along with verification of ROC curves. Based on the statistics for the number of collected LiDAR points from each object, the averag e vehicle detection accuracy is always over 80%, even though there are only less than 5 points refl ected from the testing object. In addition, the probabilistic-based system performance could be easily promoted if the amount of vehicle sources and the variety of occluded scenarios could be increased in the learning phase. The potential applications for this work include many fields such as military su rveillance, homeland security, global warming, disaster rescue, emergency road service, and criminal searching.

PAGE 111

111 Table 7-1. Summary of some morphological filtering operations. Operation Equation Comments Translation B a z a w w Bz for ,Translates the origin of B to point z. Reflection B b b w w B for ˆ Reflects all elements of B about the origin of this set. Dilation A B z B Az) ˆ ( {} “Expands” the boundary of A Erosion A A B z Bz ) ( “Contracts” the boundary of A Opening A B A ( B B ) Smoothes contours, breaks narrow isthmuses, and eliminates small islands and sharp peaks. Closing ) ( B A B A B Smoothes contours, fuses narrow breaks and long thin gulfs, and eliminates small holes. Table 7-2. Vehicle detection cross matrix Classified Vehicles Classified Clutters Reference Vehicles True Positive (TP) False Negative (FN)/Type I Error Reference Clutters False Positive (FP)/Type II Error True Negative (TN) Table 7-3. Vehicle detection accuracy for Hogtow n parking site with different LiDAR scans and overlapped all LiDAR scans. LiDAR TP FN FP TN SumAccuracy Scan #1 6 2 15 91 11485.1% Scan #2 6 2 17 92 11783.8% Scan #3 6 2 14 89 11185.6% Scan #4 5 3 9 77 9487.2% Scan #5 7 1 15 113 13688.2% Scan #6 7 1 16 93 11785.5% Scan #7 7 1 9 68 8588.2% All Scans 8 0 8 157 17395.4% Overall 52 12 103 780 94787.9% Table 7-4. Vehicle detection accuracy for Hogtown forest site with different LiDAR scans and overlapped all LiDAR scans. LiDAR TP FN FP TN SumAccuracy Scan #1 0 1 24 76 10175.3% Scan #2 0 1 13 83 9785.6% Scan #3 0 1 18 70 8978.7% Scan #4 0 1 11 84 9687.5% Scan #5 0 1 19 64 8476.2% Scan #6 1 0 16 78 9583.2% All Scans 1 0 29 119 14980.5% Overall 2 5 130 574 71181.0%

PAGE 112

112 Table 7-5. Vehicle detection accuracy for Hogtown residential site with different LiDAR scans and overlapped all LiDAR scans. LiDAR TP FN FP TN SumAccuracy Scan #1 5 1 9 29 4477.3% Scan #2 5 1 10 43 5981.4% Scan #3 4 2 7 41 5483.3% Scan #4 3 3 12 34 5271.2% Scan #5 3 3 14 48 6875.0% Scan #6 3 3 11 47 6478.1% All Scans 6 0 4 65 7594.7% Overall 29 13 67 307 41680.8%

PAGE 113

113 Figure 7-1. Object segmentation flowchart. Figure 7-2. Object segmentation result in Hogtown parking site. Figure 7-3. Object segmentation result in Hogtown forest site.

PAGE 114

114 Figure 7-4. Object segmentation result in the residential site. Figure 7-5. Object classification flowchart. Clustered Object Points Principal Component Analysis Bayesian Decision Spin Image Clutter Object Target Object Relative Entropy

PAGE 115

115 Figure 7-6. Using PCA to get the best fitting plane where the first two principal components define vectors that form a basis for the plane and the third principal component is orthogonal to the first two, and it defines the normal vector of the plane. A B Figure 7-7. An unoccluded vehicle consisting of 2 07 LiDAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-8. An occluded vehicle consisting of 82 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. -3 -2 -1 0 1 2 3 -2 0 2 -3 -2 -1 0 1 2 3 x yz -3 -2 -1 0 1 2 3 -2 0 2 -3 -2 -1 0 1 2 3 x yz

PAGE 116

116 A B Figure 7-9. An occluded vehicle consisting of 73 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-10. An occluded vehicle consisting of 41 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-11. An occluded vehicle consisting of 23 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-12. An occluded vehicle consisting of 14 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image.

PAGE 117

117 A B Figure 7-13. An occluded vehicle consisting of 12 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-14. An occluded vehicle consisting of 9 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-15. An occluded vehicle consisting of 8 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image. A B Figure 7-16. An occluded vehicle consisting of 7 Li DAR points represented in the A) 3-D shape and B) 2-D Spin image.

PAGE 118

118 Figure 7-17. Bivariate PDF of reference vehicles Figure 7-18. Bivariate PDF of reference non-vehicles Figure 7-19. The information divergence KLD for vehicles and non-vehicles

PAGE 119

119 Figure 7-20. Bivariate PDF of detected vehicles with BT =0.7 Figure 7-21. Bivariate PDF of detected non-vehicles with BT =0.7 A B Figure 7-22. Occluded vehicle detection resu lts in the Hogtown parking site from A) overlapped-scanned LiDAR data B) single-sca nned LiDAR data, where the white/black lines represent detected vehicles/non-vehicles.

PAGE 120

120 A B Figure 7-23. Occluded vehicle detection results in the Hogtown forest site from A) overlapped-scanned LiDAR data B) single-scanned LiDAR data, where the white/black lines represent detected vehicles/non-vehicles. A B Figure 7-24. Occluded vehicle detection results in the Hogtown residential site from A) overlapped-scanned LiDAR data B) single-sca nned LiDAR data, where the white/black lines represent detected vehicles/non-vehicles. Figure 7-25. Vehicle detection performance, TP, FN FP, and TN, vs. the number of collected LiDAR points from each object with the Bayesian threshold equal to 0.7.

PAGE 121

121 Figure 7-26. Vehicle detection accuracy, hit ratio an d false alarm, vs. the number of collected LiDAR points from each object with the Bayesian threshold equal to 0.7. Figure 7-27. ROC curves of our detection performance for 0, 5, 10, 15, and 20 sampling points on an object respectively. Those 6 Bayesian thresholds 0.5, 0.6, 0.7, 0.8, 0.9 and 1.5 of each curve are marked in ascending order.

PAGE 122

122CHAPTER 8 DOWNSAMPLED VEHICLE DETECTION SIMULATION Vehicle and Clutter Dataset There are 580 independent vehicles in open area which are extracted from UF campus site. The point density histogram of vehicles is shown as Figur e 8-1. The maximum length, width, and height of vehicles are about 6.3m, 2.5m and 2.3m, respectively. An envelope box is defined as a simple classifier whose size is the same as the maximum length, width, and height of vehicles. In order to increase the difficulty of the vehicle detection and balance th e total number of vehicles and clutters, the other 580 independent non-vehicle objects are selected by extracting LiDAR points inside the envelope box randomly moving in this study site. The confusion table of this envelope box classifier is in Table 8-1, which shows a simple classifier is unable to give goo d vehicle detection in this case since its recognition rate and Kappa index of agreement are 50% and 0%, respectively. Support Vector Machine The optimal Bayesian classifier is based on the es timation of the PDF functions describing the data distribution in each class. General speaking, it is a di fficult task to get a accurate distribution estimation, especially in high-dimensional spaces. Alternatively, one may make the problem easier by designing a decision surface which separates the classes from the tr aining data set such as support vector machine (SVM) [86, 87] and minimum mean square errors (MMSE), without having to deduce it from the PDFs. Although the solution may not correspond to the optimal Bayesian classifier, it usually turns out to result in better performance compared to that of the Bay es classifier which employs estimates of the involved PDFs where the size of the available training data set is limited. The general mathematical formulation of SVMs is briefly recalled as follows. Linear SVM Given some training data D and label space Y (e.g., } 1 1 { Y R Dn in a two-class problem). The classification is carried out using a linear discriminant function ) ( Y D Each D xi is a

PAGE 123

123 n-dimensional real vector which is an available training sample with a label iy, where ] 1 [ N i The theoretical aim of supervised classification is to fi nd the maximum-margin hyperplane that divides the points having 1 iy from those having 1 iy. For a linear classifier, 0w x w x where D w is the normal vector to the hyperplane and 0w is the bias. We aim at finding the classifier parameters ) (0w w which verify: 0 ) ( ) (0 w x w y D y xi i i i (8.1) Since the SVM method searches the best classifier (i.e., the largest margin), we impose: 1 ) ( ) (0 w x w y D y xi i i i (8.2) The support vectors lie on two hyperplanes 10 w x w which are parallel and equidistant to the optimal linear separable hyperplane. Finally, the optimal hyperplane has to maximize the margin (i.e., the Euclidian distance between both hyperplanes, defined as w2) under the constraints defined in Equation (8.2). Unfortunately, in most cases, such quadratic optimization problems are unsolvable: we cannot find a linear classifier consistent with the training set. The classification problem is not linearly separable. Consequently, slack variables or margin errors i where a slack variable is a nonnegative variable that turns an inequality into an equality constraint, are introduced to cope with misclassified samples and prevent Equation (8.2) from being violated. Another r eason is the avoidance of over-fitting the classifier to the training samples, which would result in poor performance. It becomes: i i i i iw x w y D y x 1 ) ( ) (0, (8.3) where 0 1 iN i The final optimization problem is subsequently: N i iC w1 22 min subject to Equation (8.3) (8.4) C is a constant which determines the trade-off between margin maximization and training error

PAGE 124

124 minimization. It turns out that the solution is gi ven as a weighted average of the training points: N i i i ix y w1 (8.5) The coefficients i are the Lagrange multipliers of the optimization task and they are zero for all points outside the margin and on the correct side of the classifier. These points therefore do not contribute to the formation of the direction of the classifier. The rest of the points, with nonzero i ’s, which contribute to the buildup of w are called support vectors. Multi-Class SVM SVMs are designed to solve binary problems. In the cases of having more than two classes of interest, the dominating approach for doing so is to reduce the single multiclass problem into multiple binary classification problems. For such pairwise classification, 2 ) 1 ( n n binary classifiers are computed on each pair of classes. Each of the problem s yields a binary classifier, which is assumed to produce an output function that gives relatively large values for examples from the positive class and relatively small values for examples belonging to the negative class. Each sample is assigned to the class getting the highest number of votes. A vote for a given class is defined as a classifier assigning the sample to that class. Nonlinear SVM When the classification problem is not linearly sep arable, one solution consists in changing the feature space. We can create nonlinear classifiers by applying the kernel trick to maximum-margin hyperplanes. The resulting algorithm is formally simi lar, except that every dot product is replaced by a nonlinear kernel function ) ( ), ( ) (j i j ix x x x K where denotes the inner product operation in a higher dimension space. This allows the algorithm to fit the maximum-margin hyperplane in a transformed feature space. Thus, to solve a lin ear problem in the high-dimensional space, all we have to do is replace the inner products with the corresponding kernel evaluations. Typical examples of kernel functions are polynomial function and the radial basis function (RBF), defined as

PAGE 125

125 n Ty x y x K (8.6) and 2 2exp ,y x y x K, (8.7) respectively, where and n are user-defined parameters in Equation (8.6) and is a user-defined parameter that specifies the rate of decay of y x K toward zero, as y moves away from x in Equation (8.7). Note that solving a linear problem in th e high-dimensional space is equivalent to solving a nonlinear problem in the original space. As in Equa tion (8.5), the hyperplane computed by the SVM method in the high-dimensional space is N i i i ix y w1) ( (8.8) Given an x we first map it to ) ( x and then test whether the following is less than or greater than zero: 0 1 0 1 0) ( ) ( ), ( w x x K y w x x y w x w x gN i i i i N i i i i (8.9) From the previous relation, it becomes clear th at the explicit form of the mapping function ) ( is not required; all we have to know is the kernel func tion since data appear only in inner products. Observe that the resulting discriminant function, ) ( x g, is nonlinear because of the nonlinearity of the kernel function. For the kernel function selection, the radial basis function was selected in this study since its detection accuracy is little better than the polynomial function. Novel Feature Extraction Spin Image Features Two Spin image features and developed in Chapter 7 are used here again. One vehicle

PAGE 126

126 DSM and one clutter DSM with their LiDAR points are showed in Figure 8-2, for illustration. The obtained Spin images for the vehicle and clutter example are shown in Figure 8-3. It can be found that the highest counting positions in each column of the vehicle are similar and less varied than those of the clutter which give the ability to discr iminate between vehicles and clutters. If we only use these two Spin image features and as the whole feature space in the SVM classifier, the classification result can be obtained as Figure 8-4 and Table 8-2. The recognition rate and Kappa index of agreement by using and in the SVM are 99.66% and 99.31%, respectively, which are almost perfect and much better than the simple envelope box classifier. Principal Component Features Another piece of information that is useful to distinguish vehicles from clutter is the blocking LiDAR area. A vehicle is a solid object which cannot be penetrated by LiDAR scanning, so no LiDAR points exist in its underneath area. For a non-solid clutter such as a tree, some ground points could be found inside the area it occupies. The downsampling situation makes this feature more significant since the blocking LiDAR area will not be reduced for a solid object. However, if we only use collected LiDAR points from an object, its occupied area could be reduced a lot in a very sparse point density situation. Therefore, not only object points but also ground points are considered for computing a blocking area. The blocking area by a vehicle tends to be a rectangular shape whose length and width can be estimated by principal component analysis in Chapter 7. The length and width estimation for the vehicle and cl utter example is showed in Figure 8-5. If we only use these two features blocking length and width as the whole feature space in the SVM classifier, the classification result can be obtained as Figure 8-6 and Table 8-3. Its recognition rate and Kappa index of agreement are 91.98% and 83.97%, respectively, wh ich are worse than the spin image features but still much better than the simple envelope box classifier. Surface Intensity Feature Return intensity or simply intensity is an a ttribute that describes the strength of the beam

PAGE 127

127 backscatter pertaining to the return in question. It de pends on the reflectance properties of the target, and hence it can potentially be used in target discrimin ation. Its utility for object classification is often reduced because of its dependence on bidirectional re flectance distribution function effects, the distance (range) to the laser instrument, the total number of re turns identified in the parent beam, the rank of the return (first, second, etc.) in the parent beam, and the r eceiver’s gain factor [88]. In this study, we still try to take advantage of intensity information to extract a useful feature. The intensity map is generated by gridding x, y, and intensity data with triangle-based linear interpolati on (Figure 8-7). It is observed that intensity of vehicles circled by white lines vari ed not only between different vehicles but also on individual vehicles. It seems that the LiDAR intensity needs to be normalized before using. So, we also study a paper [89] for how to get LiDAR intensity normalization. Concerning full-waveform laser data for each single beam, the total number of detected backscattered pulses is known and is assigned to the corresponding echoes. Each echo is described by a point with its 3D coordinate, signal amplitude a, and signal width w at full-width-at-half-maximum (FWHM) derived from the Gaussian approxima tion (such as Gaussian Mixture Model: GMM). Additionally the 3D coordi nate of the sensor position is available. The shape of the received waveform depends on the illuminated surface area, especially on the material, reflectance of the surface and the inclination angle between the surface normal and the laser beam direction. For all points with high planarity, the measured intensity can be normalized by ) cos( RI I where RIis recorded intensity and ) cos( is the incidence angle [89]. The normalized intensity can be obtained if 1) the flight position corresponding to each LiDAR point is given. 2) LiDAR points must be located on hi gh planarity to get an accurate normal vector of the reflected surface of a target, which depends on the relative roughness of the surface. When the LiDAR point density is high enough, the normal vector can be calculated accurately even if the size of the flat plane part is small. But, if the LiDAR point density is too low, the normal vector will not be accurate if the size of the flat plane part is not large enough, such as a vehicle. Instead of normalized intensity, I come up with a no vel idea, surface intensity index (SII). First, we

PAGE 128

128 have to find the 3D convex hull from those points belonging to one object iS, ] 1 [ n i The total point number of iS is represented by iS N. It is assumed that those points, ] 1 [ :i ijS N j p of iS on the convex hull reflect all measured energy or intensity, ijI, to its surface. The rest points inside the convex hull can only reflect partial measured energy to its surface which is defined by the power propagation rule from wireless communications n ij ij ijd d I I 0 (8.10) where 0d is a reference distance, ijd is the shortest distance from the point ijpto the convex hull of iS and n is the path loss exponent. In this study, I assume 3 00 dm and 2 n if 0d dij, otherwise ij ijI I The SII of iSis between 0 and 1, which is defined as i iS N j ij S N j ij iI I SII1 1 (8.11) The SIIs for the vehicle and clutter example ar e shown in Figure 8-8. If we use this feature combined with blocking area of an object in the SV M classifier, the classification result can be obtained as Figure 8-9 and Table 8-4. Its recognition rate a nd Kappa index of agreement are 87.59% and 75.17%, respectively, which are worse than our previous feat ures but still much better than the simple envelope box classifier. Vehicle Detection Methods Vehicle Recognition #1 Method In [90], the authors used a six-parameter representa tion that includes the vehicle length, width, and four vehicle parameters (average height values, h1-h4, computed over the four equally size regions) as shown in Figure 8-10. 72 vehicles were chosen a nd processed in an interactive way, the regions containing vehicles were selected by an operator a nd the vehicles were automatically extracted by the height threshold method. Their vehicles were paramete rized and then categorized into three main groups:

PAGE 129

129 passenger cars, multi-purpose vehicles such as SUVs, minivans, light trucks, and trucks/eighteen-wheelers. To reduce the dimensionality of the parameter sp ace, PCA was then performed. PCA is an effective tool for handling data representation or classificati on problems where there is a significant correlation among the parameters describing the object patterns. By training the datasets, the correlation can be determined and a reduced parameter set can be define d that can both represent the information in a more compact way and can support an efficient classificati on in the reduced feature space. The clear advantage of the method is that it does not require any physical modeling of the data; of course, the selection of the input parameters has some importance. Provided that a rich set of input parameters is defined, the method will effectively identify the redundancy and t hus usually results in a quite reduced parameter representation. In their investigations, the 72 vehicles provided a statistically meaningful dataset for the PCA process. Although the six-parameter representation in our 580 vehicles and 580 clutters do not provide a dataset as meaningful as their investigations, the fi rst two principal components of this feature space still represent about 85% of the information (Figure 8-11). In order to convey as much information as possi ble through the PCA process to the classifier, we choose those features which represent at least 99% of the information in the PCA. Therefore, all six principal components are chosen and used in the SV M classifier. Its recognition rate and Kappa index of agreement are 87.59% and 75.17%, r espectively, shown in Table 8-5. Vehicle Recognition #2 Method Applying more features could improve the perfo rmance of PCA-based classification. Since the LiDAR dataset also contains intensity values, the aut hors in [91] use them as additional parameters. They supposed different vehicle categories produce differen t reflection intensity and extend the six-parameter representation to ten-parameter representation by adding four mean intensity values (int1-int4) corresponding to four mean heights (h1-h4). As op posed to the strikingly positive results of the PCA method which is based on geometric parameters, this enhanced algorithm did not result in well

PAGE 130

130 distinguished categories; the deviation between th e intensity values turned out to be too high. In this study, we still try to use this ten-pa rameter representation to see its performance. The PCA process reduces its redundancy from 10 features to 4 principal components which can represent over 99% of the information of original features (Figur e 8-12). However, using these 4 principal components in the SVM classifier does not result in a satisfactor y outcome. Its recognition rate and Kappa index of agreement are only 67.50% and 35.00%, respectively, shown in Table 8-6. SVM with SPI (Spin image, PCA, and Intensity) Features A Spin image, PCA, and surface Intensity inde x (SPI) method is proposed here, which is a five-parameter representation including the aforementioned ,, blocking length, blocking width, and SII. The PCA process reduces its redundancy from 5 features to 4 principal components which can represent about 99.9% of the information of original features (Figure 8-13). Applying these 4 principal components to the SVM classifier, it results in a very satisfactory performance. Its recognition rate and Kappa index of agreement are only 99.91% and 99.83%, respectively, shown in Table 8-7. Downsampled Vehicle Detection Test For those 580 unoccluded vehicles, we generate th e test datasets by using downsampling rates from 0.1 to 0.9 with an interval of 0.1. Since the other 580 clutters are from diverse objects occluded by trees irregularly, they are not downsampled again in the t est datasets. For each downsampling rate, there are 10 simulations in which vehicle points are randomly sampled. The true positive and false negative values are averaged by 10 simulations to get the average recognition rate a nd Kappa index of agreement in each downsampling rate. Tables 8-8 to 8-10 show th e downsampled vehicle detection simulation performances for the vehicle recognition #1, vehicl e recognition #2, and SPI methods. Figure 8-14 and Figure 8-15 show the comparison of the envelope box method and the above three methods in average recognition rate and average Kappa index of ag reement for the downsampled vehicle detection simulation. It is obvious by looking at Figure 8-15 that our proposed SPI method is the best one compared to the other three methods, especially in low sampling ra tes. That is because we tr y to mitigate the impact

PAGE 131

131 of sparse samples on vehicle detection during the feat ure extraction stage. Figure 8-15 also verifies that the five features in the SPI method can allow more ser ious under-sampling situa tions and tolerate more shape distortion errors than others. Testing SVM with SPI Features on Hogtown forest sites The above occluded vehicle detection results are obtained from downsampling simulations for vehicles in open areas, a scenario which could be diff erent from the real-world situation. Thus, we apply our proposed SPI method to three forested sites in Hogtown area as well, where the training data are still from Hogtown parking site and Hogtow n forest site. The five features extracted from Spin image, PCA, and LiDAR intensity in Hogtown area still can be re duced to four principal components (Figure 8-16) even though the variation percentages of them are different to UF campus area. The occluded vehicle detection results for Hogtown parking site, Hogtown fo rest site, and Hogtown r esidential site are shown in Tables 8-11, 8-12 and 8-13, respectively. Comparing the detection results between the methods based on Bayesian decision with Spin image features and SVM classifier with SPI features by Tables 7-3, 7-4, 7-5, 8-11, 8-12 and 8-13, shows that the latter method improves the performance of the former method. The overall performances of the former one for Hogtown parking site, Hogt own forest site, and Hogtown resi dential site are 87.9%, 81.0%, and 80.8%, respectively, while they are promoted by the latter method to 97.15%, 99.16%, and 93.03%, respectively. It is noted that the vehicle missing errors in these sites become larger because the SVM classifier focuses on the overall performance. If some one needs to balance the missing error and false alarm of vehicle detection, our Bayesian decision c onsidering with relative entropy values for the vehicle class and non-vehicle class is a good choice since it will give a tradeoff between vehicle and non-vehicle decisions. However, the vehicle missing errors in th e SVM classifier with SPI feature can be easily decreased if the number of independent vehicl es in occluded situations can be increased. Summary The proposed SPI method is a new vehicle detection approach which combines five features extracted from Spin image, PCA, and LiDAR intensity and applies them to the SVM classifier. The main

PAGE 132

132 advantage for these features is that they can mitigat e the impact of sparse samples and tolerate more shape distortion errors. By using the independent 580 vehicles and 580 non-vehicle objects in the dataset, it is verified that this SPI method outperforms the othe r three methods for vehicle detection, especially in low sampling rates. In addition, we also apply this method to three forested sites in Hogtown area. It shows that the overall performances for those sites in Bayesian decision are improved significantly where the average accuracy of the three s ites is promoted from 83.23% to 96.45%.

PAGE 133

133 Table 8-1. Confusion table of the envelope box classifier Envelope Box Classified VehiclesClassified Clutters Reference Vehicles 5800 Reference Clutters 5800 (Recognition rate, Kappa index of agreement) = (50%, 0%) Table 8-2. Confusion table of the SVM classification by spin-image features Spin Image ( ,) Classified VehiclesClassified Clutters Reference Vehicles 5791 Reference Clutters 3577 (Recognition rate, Kappa index of agreement) = (99.66%, 99.31%) Table 8-3. Confusion table of the SVM classificati on by features (blocking length, blocking width) PCA (Length, Width) Classified VehiclesClassified Clutters Reference Vehicles 55525 Reference Clutters 68512 (Recognition rate, Kappa index of agreement) = (91.98%, 83.97%) Table 8-4. Confusion table of the SVM classification by features (SII, blocking area) (SII, blocking area) Classified VehiclesClassified Clutters Reference Vehicles 55030 Reference Clutters 114466 (Recognition rate, Kappa index of agreement) = (87.59%, 75.17%) Table 8-5. Confusion table of the SVM classification by features of the vehicle recognition #1 Vehicle Recognition #1 Classified VehiclesClassified Clutters Reference Vehicles 5728 Reference Clutters 12568 (Recognition rate, Kappa index of agreement) = (98.28%, 96.55%) Table 8-6. Confusion table of the SVM classification by features of the vehicle recognition #2 Vehicle Recognition #2 Classified VehiclesClassified Clutters Reference Vehicles 346234 Reference Clutters 143437 (Recognition rate, Kappa index of agreement) = (67.50%, 35.00%) Table 8-7. Confusion table of the SVM classification by features of SPI method SPI Classified VehiclesClassified Clutters Reference Vehicles 5800 Reference Clutters 1579 (Recognition rate, Kappa index of agreement) = (99.91%, 99.83%)

PAGE 134

134 Table 8-8. Average vehicle detection perform ance of Vehicle Recognition #1 method with 10 simulations for each downsample rate Downsample Rate True Positive Fal se NegativeRecognition RateKappa Index 0.1 43.80 536.2052.74%5.48% 0.2 248.10 331.9070.35%40.71% 0.3 398.00 182.0083.28%66.55% 0.4 474.60 105.4089.88%79.76% 0.5 511.00 69.0093.02%86.03% 0.6 528.00 52.0094.48%88.97% 0.7 537.70 42.3095.32%90.64% 0.8 545.10 34.9095.96%91.91% 0.9 548.10 31.9096.22%92.43% 1.0 572.00 8.0098.28%96.55% (False Positive, True Negative) = (12, 568) Table 8-9. Average vehicle detection perform ance of Vehicle Recognition #2 method with 10 simulations for each downsample rate Downsample Rate True Positive Fal se NegativeRecognition RateKappa Index 0.1 232.50 347.5057.72%15.43% 0.2 253.20 326.8059.50%19.00% 0.3 265.60 314.4060.57%21.14% 0.4 277.40 302.6061.59%23.17% 0.5 300.70 279.3063.59%27.19% 0.6 306.10 273.9064.06%28.12% 0.7 315.60 264.4064.88%29.76% 0.8 315.20 264.8064.84%29.69% 0.9 315.90 264.1064.91%29.81% 1.0 346.00 234.0067.50%35.00% (False Positive, True Negative) = (143, 437) Table 8-10. Average vehicle detection performance of SPI method with 10 simulations for each downsample rate Downsample Rate True Positive Fal se NegativeRecognition RateKappa Index 0.1 449.20 130.8088.64%77.28% 0.2 529.40 50.6095.55%91.10% 0.3 557.60 22.4097.98%95.97% 0.4 565.60 14.4098.67%97.34% 0.5 571.20 8.8099.16%98.31% 0.6 572.00 8.0099.22%98.45% 0.7 572.80 7.2099.29%98.59% 0.8 573.30 6.7099.34%98.67% 0.9 573.70 6.3099.37%98.74% 1.0 580.00 0.0099.91%99.83% (False Positive, True Negative) = (1, 579)

PAGE 135

135 Table 8-11. Vehicle detection accuracy for Hogtow n parking site with different LiDAR scans and overlapped all LiDAR scans based on SVM with SPI features. LiDAR TP FN FP TN SumAccuracy Scan #1 2 6 0 106 11494.74% Scan #2 3 5 0 109 11795.73% Scan #3 3 5 0 103 11195.50% Scan #4 5 3 0 86 9496.81% Scan #5 6 2 0 128 13698.53% Scan #6 5 3 0 109 11797.44% Scan #7 8 0 0 77 85100.00% All Scans 5 3 0 165 17398.27% Overall 37 27 0 883 94797.15% Table 8-12. Vehicle detection accuracy for Hogtown forest site with different LiDAR scans and overlapped all LiDAR scans based on SVM with SPI features. LiDAR TP FN FP TN SumAccuracy Scan #1 0 1 0 100 10199.01% Scan #2 0 1 0 96 9798.97% Scan #3 0 1 0 88 8998.88% Scan #4 1 0 0 95 96100.00% Scan #5 0 1 0 83 8498.81% Scan #6 0 1 0 94 9598.95% All Scans 0 1 0 148 14999.33% Overall 1 6 0 704 71199.16% Table 8-13. Vehicle detection accuracy for Hogtown residential site with different LiDAR scans and overlapped all LiDAR scans. LiDAR TP FN FP TN SumAccuracy Scan #1 2 4 0 38 4490.91% Scan #2 1 5 0 53 5991.53% Scan #3 1 5 0 48 5490.74% Scan #4 2 4 0 46 5292.31% Scan #5 2 4 0 62 6894.12% Scan #6 2 4 1 57 6492.19% All Scans 4 2 0 69 7597.33% Overall 14 28 1 373 41693.03%

PAGE 136

136 2.5 3 3.5 4 4.5 5 5.5 6 6.5 0 20 40 60 80 100 120 #LiDAR points/m2 per vehicleFreq Figure 8-1. LiDAR point density histogram of reference vehicles in the UF campus site, where 5 4 and 6 0 x (meter)y (meter) 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 0.2 0.4 0.6 0.8 1 1.2 1.4 A x (meter)y (meter) 1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8 9 10 11 12 13 0 0.5 1 1.5 2 B Figure 8-2. A DSM and LiDAR point map example for A) a vehicle and B) a clutter. x (meter)y (meter) 0.4 0.8 1.2 1.6 2 2.4 2.8 0.2 0.4 0.6 0.8 2 4 6 8 10 A x (meter)y (meter) 0.4 0.8 1.2 1.6 2 2.4 2.8 3.2 3.6 4 4.4 0.4 0.8 1.2 1.6 2 2.4 2.8 0 5 10 15 20 B Figure 8-3. Spin image of A) a vehicle and B) a clutter in Figure 8-2a and 8-2b, respectively.

PAGE 137

137 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5 3 3.5 Object Feature #1: Object Feature #2: Vehicle Clutter Figure 8-4. SVM classification result by Spin-image features ) ( x (meter)y (meter) 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 0.2 0.4 0.6 0.8 1 1.2 1.4 A x (meter)y (meter) 1 2 3 4 5 6 7 8 1 2 3 4 5 0 0.5 1 1.5 2 B Figure 8-5. Length and width estimation by PCA for A) a vehicle and B) a clutter in Figure 8-2a and 8-2b, respectively.

PAGE 138

138 0 2 4 6 8 10 0 1 2 3 4 5 6 Object Feature #3: LengthObject Feature #4: Width Vehicle Clutter Figure 8-6. SVM classification result by blocking length and blocking width features. x (meter)y (meter) 25 50 75 100 125 150 25 50 75 100 125 150 0 50 100 150 200 250 300 Figure 8-7. Intensity Map gridded from LiDAR da ta (x,y,int) at one of the UF campus site. x (meter)y (meter) 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 0.2 0.4 0.6 0.8 1 1.2 1.4 Points on Convex Hull Points Inside Convex Hull A x (meter)y (meter) 1 2 3 4 5 6 7 8 1 2 3 4 5 0 0.5 1 1.5 2 Points on Convex Hull Points Inside Convex Hull B Figure 8-8. Point relationship with the convex hull fo r A) a vehicle and B) a clutter in Figure 8-2, where their surface intensity index are 100% and 94%, respectively.

PAGE 139

139 0.86 0.88 0.9 0.92 0.94 0.96 0.98 1 0 5 10 15 20 25 30 35 40 Object Feature #5: Surface Intensity Index (SII)Object Feature #6: Area y() Vehicle Clutter Figure 8-9. SVM classification result by surf ace intensity index and blocking area features. Figure 8-10. The six-parameter representation (W, L, H1, H2, H3, H4) of the vehicle recognition #1 method.

PAGE 140

140 1 2 3 4 5 6 0 10 20 30 40 50 60 70 80 90 100 Principal ComponentVariance Explained (%)70.57% 14.25% 4.99% 4.11% 3.91% 2.16% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Figure 8-11. Contained information of PCA of the vehicle recognition #1 method. 1 2 3 4 5 6 7 8 9 10 0 10 20 30 40 50 60 70 80 90 100 Principal ComponentVariance Explained (%)53.13% 21.14% 15.86% 9.84% 0.02% 0.00% 0.00% 0.00% 0.00% 0.00% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Figure 8-12. Contained information of PCA of the vehicle recognition #2 method. 1 2 3 4 5 0 10 20 30 40 50 60 70 80 90 Principal ComponentVariance Explained (%)78.27% 14.02% 4.70% 3.01% 0.01% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% Figure 8-13. Contained information of PCA of the novel features from Spin image, PCA, and LiDAR intensity (SPI).

PAGE 141

141 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Downsample RateRecognition Rate Envelope Box Vehicle Recognition #1 Vehicle Recognition #2 SPI Algorithm Figure 8-14. Average recognition rate of vehicle detection vs. downsample rate comparison among envelop box, vehicle recognition #1, ve hicle recognition #2, and SPI algorithm. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Downsample RateKappa Index of Agreement Envelope Box Vehicle Recognition #1 Vehicle Recognition #2 SPI Algorithm Figure 8-15. Average Kappa index of agreement of vehicle detection vs. downsample rate comparison among envelop box, vehicle recognition #1, ve hicle recognition #2, and SPI algorithm.

PAGE 142

142 1 2 3 4 5 0 10 20 30 40 50 60 70 80 90 100 Principal ComponentVariance Explained (%)42.77% 27.35% 26.41% 2.87% 0.60% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Figure 8-16. Contained information of principal components of the SPI method applied to Hogtown forest sites.

PAGE 143

143CHAPTER 9 CONCLUSIONS AND CONTRIBUTIONS Conclusions Tree Canopy Removal ISPRS Working Group III/3 conducted a test and foun d that all bare ground point extraction filter algorithms perform well on LiDAR point clouds from sm ooth rural landscapes, but all produce errors in rough terrain with vegetation canopy. We develop this canopy removal algorithm to help detect obscure targets underneath forest canopy as well as mitigat e the vegetation problem for those filters. In this algorithm, the multiple-return characteristi c of LiDAR data is analyzed and accordingly laser shots are classified as single-return or multiple-return shots. The major challenge of removing canopy is that some foliage will unexpectedly reflect single-re turn shots rather than normal multiple-return shots when they are very dense. This challenge can be solved by using our developed algorithms such as analyzing distance relationships between foliage applying morphological filters to process the canopy/non-canopy image and creating rough digital terra in models to calculate above ground levels of points, etc. The unique feature in this algorithm is that no parameter tweaking is required. Both of the city and forest sites are tested where the data are from ISPRS and UF, respectively. It shows that all tree or forest canopy points have been removed such th at all obscure vehicles or buildings underneath canopy can now be easily seen. Bare-Earth Extraction A DTM, commonly used interchangeably with a DEM, is a digital representation consisting of terrain elevations for ground positions. A DTM is also called a bare-Earth model since it excludes features on the Earth such as tall vegetation, buildings, and bridges. Thus, a DTM can be used to generate graphics displaying terrain slope, direction of slope, and terrain profiles. As such it is typically applied to flood risk assessment, drainage modeli ng, multitemporal land use and land cover changes, landslide and mudslide monitoring, as we ll as urban planning and management.

PAGE 144

144 So far, more prevalent methods of LiDAR f iltering can be categorized into three groups: morphological filtering, segmentation modeling, and su rface modeling. In this work, we develop a novel bare-earth extraction algorithm consisting of seg mentation modeling and surface modeling based on the tree canopy removal algorithm including the morphological filtering. The proposed segmentation modeling is built on a triangulated irregular netw ork and composed of triangle assimilation, edge clustering, and point classification to achieve be tter discrimination of objects and preserve terrain discontinuities. The surface modeling is proposed to it eratively correct both Type I and Type II errors through estimating roughness of dig ital surface/terrain models, detecting bridges and sharp ridges, etc. This proposed method is compared with twelve ot her filters working on the same fifteen study sites provided by ISPRS. Our average error and kappa inde x of agreement in the automated process are 4.6% and 84.5%, respectively, which outperform all the othe r twelve proposed filters. Our kappa index, 84.5%, can be interpreted as almost perfect agreement. In addition, applying this work with optimized parameters further improves performance. Recently, an unsupervised classification algorithm called Skewness Balancing was developed for object and ground point separation in airborne Li DAR data. Although the main advantages of their algorithm are threshold-freedom and independence from LiDAR data resolution, they have to build a prediction model to categorize LiDAR tiles as hilly or moderate terrains. Howeve r, not all LiDAR data can be categorized as either completely hilly or m oderate terrain tiles. Once a tile includes both terrain types, their algorithm will face a big challenge. Insp ired by their algorithm, we develop a novel slope-based statistical algorithm which is appropriate to any mixed or complicated terrain types. Initially, most objects are removed and initial terrains can be obtained in our object detection algorithm. Slope differences can be assumed to be a zero-mean normal dist ribution in all kinds of terrains, unlike absolute height information used by the Skewness Balanci ng algorithm. Based on slope difference variations, the Chi distribution measurement is used to decide the adaptive slope threshold. Accordingly, the adaptive growing height threshold of each pixel is derived by 8-connected neighbor pixels which can be used to iteratively correct classified points in the initial terrain The testing results show that this algorithm is

PAGE 145

145 even better than our previous algorithms, Chang#1 an d Chang#2, which have outperformed all twelve other algorithms working on the same study sites. Occluded Vehicle Detection Vehicle detection research is devoted to the Inte lligent Transportation System (ITS) and Automatic Vehicle Guidance (AVG), but is seldom exploited in th e forested terrain. The state-of-the-art airborne LiDAR can provide data in large sp atial extents with varying temporal resolution and can be deployed more or less anywhere and at any time which can be potentially applied to for ested terrain for military surveillance, homeland security, global warming, disast er rescue, emergency road service, and criminal searching. In this work, under-canopy LiDAR points are obtained and clustered by the canopy removal algorithm, bare-Earth extraction, and morphological image processing. Clustered objects are classified into the vehicle or non-vehicle class by Bayesian decision based on Spin images forming the feature space and information divergences determining the optim al Bayesian threshold. The vehicle detection results are demonstrated, discussed, and verified by Receiver Operating Characteristic curves, where diverse-scanned accuracies of training and testing sites range between 73% and 95%. In addition, we propose the SPI method, a novel occluded vehicle detection approach, which combines five features extracted from Spin image, PC A, and LiDAR intensity and applies them to the SVM classifier. This SPI method is compared to a simple method and two other vehicle detection methods proposed by other authors’ papers published in ISPRS. With ten simulations in each different downsampling rate testing on inde pendent 580 vehicles and 580 non-vehicle objects, our experiments show that this SPI method outperforms all other methods, especially in low sampling rates. Contributions Although the vehicle detection has been devoted in the Intelligent Transportation System (ITS), Automatic Vehicle Guidance (AVG), and traffic flow estimation in the LiDAR applications, it has not been exploited in the forested terrain with cluttered environment. In this study, we finished a system whose goal is to detect underneath vehicles in for ested terrain from LiDAR point clouds. However, this system can contribute in three aspects.

PAGE 146

146 In the first aspect, our canopy removal algorithm can help 1) detect obscure targets underneath forest canopy and 2) mitigate the vegetation problem for those DTM extraction algorithms. As a matter of fact, the thermal imaging cameras can see the heat signature of people, boats, and vehicles in total darkness as well as through smoke, haze, and light fog, but not through the forest canopy. Whether in a city or a forest site, the vegetation area can be co rrectly detected and canopy points are successfully removed. All obscure vehicles or buildings underneat h tree canopy are revealed as we demonstrated. Accordingly, the occluded rate of forest canopy can be obtained. Furthermore, the detailed x-y distribution of the remaining point density can be fo und as well which will be very useful for predicting the performance of occluded target detecti on with respect to various object locations. In the second aspect, our automatic and robust ground filtering is important in LiDAR applications where classified ground and object points can be used for DTM generation and further reconstruction of topographic features. Many methods have been proposed to extract bare-Earth points from LiDAR data. However, most filters perform well in flat a nd uncomplicated landscapes, while the landscapes containing steep slopes and discontinuities are still a problem which has not been fully solved. The performance of our point-based bare-Earth extraction algorithm has been compared to twelve proposed filters and evaluated by working on the same fifteen st udy sites. The average total errors and kappa index of agreement of this work in the automated process are 4.6% and 84.5%, respectively, which outperforms all twelve other filters and such kappa index is interpreted as almost perfect agreement. This algorithm applied with optimized parameters performs even better. In addition, another developed grid-based bare-Earth extraction is a slope-based statistical algor ithm which can be adaptive with site-wide and local height variations by the Chi distribution measuremen t and the derived flexible height threshold. This novel algorithm further improves the ground point extr action performance, with average total errors and kappa index of agreement of 3.4% and 88.6%, respectively. In the last aspect, we demonstrate that the stat e-of-the-art airborne LiDAR system can provide valuable data which can effectively support the occluded vehicle detection in forest terrain. Based on the actual number of collected LiDAR points from each object, the average vehicle detection accuracy is

PAGE 147

147 always over 80% by our proposed Bayesian framework w ith Spin image features, even though there are only less than 5 points reflected from the testing object. This probabilistic-based system performance could be easily promoted if the amount of vehicle sources and the variety of occluded scenarios could be increased in the learning phase. In addition, we al so propose SPI method which combines five features extracted from Spin image, PCA, and LiDAR intensity and applies them to the SVM classifier. The main advantage for these features is that they can mitigat e the impact of sparse samples and tolerate more shape distortion errors. By using the independent 580 vehicles and 580 non-vehicle objects in the dataset, it is verified that this SPI method outperforms the other three vehicle detection methods, especially in low sampling rates. The potential applications for this work include many fields such as military surveillance, homeland security, global warming, disast er rescue, emergency road service, and criminal searching for those vehicles occluded in forested terrain.

PAGE 148

148 LIST OF REFERENCES [1] K. C. Slatton, W. E. Carter, R. L. Shrestha, and W. Dietrich, “Airborne Laser Swath Mapping: Achieving the resolution and accuracy required for geosurficial research,” Geophys. Res. Lett. 34 2007. [2] W. E. Carter, R. L. Shrestha, and K. C. Slatton, “Geodetic Laser Scanning,” Physics Today 41 47, Dec (2007). [3] K. C. Slatton, M. M. Crawford, and B. L. Evans, “Fusing Interferometric Radar and Laser Altimeter Data to Estimate Surface Topography and Vegetation Heights”, IEEE Transactions on Geoscience and Remote Sensing 39 (11), 2470 2482, (2001). [4] G. Sithole and G. Vosselman, “Experimental comparison of filter algorithms for bare-Earth extraction from airborne laser scanning point clouds,” ISPRS J. Photogramm. Remote Sens ., 59 85 101, (2004). [5] K. Kampa and K. C. Slatton, “An Adaptive Multiscale Filter for Segmenting Vegetation in ALSM Data,” Proc. IEEE 2004 International Geoscien ce and Remote Sensi ng Symposium (IGARSS) 6 3837 3840, Sep. (2004). [6] H. Cho, K. Kampa, and K. C. Slatton, “M orphological Segmentation of LiDAR Digital Elevation Models to Extract Stream Channels in Forested Terrain,” Proc. IEEE 2007 International Geoscience and Remote Sensing Symposium (IGARSS) Barcelona, Spain, 23 27, 3182 3185, Jul., (2007). [7] H. Lee, K. C. Slatton, B. Roth, W. P. Cropper, “Prediction of Forest Canopy Sunlight Distribution Using Airborne LiDAR Data,” International Journal of Remote Sensing 30, 189207, (2009). [8] B. Koch, U. Heyder, and H. Weinacker, “Detec tion of individual tree crowns in airborne LiDAR data,” Photogrammetric Engineering and Remote Sensing 72 (4), 357 363, (2006). [9] H. Lee, K. C. Slatton, B. Roth, and W. P. Cropper, “Adaptive Clustering of Airborne LiDAR Data to Segment Individual Tree Crowns in Managed Pine Forests,” International Journal of Remote Sensing 31, 117139, (2010). [10] G. Sithole, and G. Vosselman, “Experimental comparison of filter algorithms for bare earth extraction from airborne laser scanning point clouds,” ISPRS Journal of Photogrammetry and Remote Sensing 59 85 101 (2004). [11] G. Vosselman, “Slope based filtering of Laser altimetry data. International Archives of Photogrammetry,” Remote Sensing and Spatial Information Sciences 33 (Part B3-2), 935–942, (2000). [12] J. Kilian, N. Haala, and M. Englich, “Captu re and evaluation of airbor ne laser scanner data,” International Archives of P hotogrammetry and Remote Sensing 31 (Part B3), 383–388, (1996). [13] K. Zhang, S. Chen, D. Whitman, M. Shyu, J. Yan, and C. Zhang, “A progressive morphological filter for removing nonground measure ments from airborne LIDAR data,” IEEE Transactions on Geoscience and Remote Sensing 41 (4), 872–882, (2003).

PAGE 149

149 [14] G. Sithole, “Filtering of laser altimetry data using slope adaptive filter,” International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences 34 (Part 3/W4), 203–210, (2001). [15] M. Roggero, “Airborne laser scanning: clustering in raw data,” International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences 34 (Part 3/W4), 227–232, (2001). [16] K. Kampa and K.C. Slatton, “An adaptive multiscale filter for segmenting vegetation in ALSM data,” IEEE International Geoscience and Re mote Sensing Symposium, IGARSS 6 3837–3840, (2004). [17] G. Sithole, and G. Vosselman, “Filtering of airborne laser scanner data based on segmented point clouds,” International Archives of Photogrammetry, Remote Sensing and Spatial Information Sciences 36 (Part 3/W19), 66–71, (2005). [18] C. Nardinocchi, G. Forlani, and P. Zingare tti, “Classification and filtering of laser data,” International Archives of P hotogrammetry and Remote Sensing 34 (Part 3/W13), Dresden, Germany, (2003). [19] S. Filin, “Surface clustering from airborne laser scanning data,” International Archives of Photogrammetry and Remote Sensing 34 (3A), Graz, Austria, (2002). [20] G. Vosselman and S. Dijkman, “3D building m odel reconstruction from poi nt clouds and ground plans,” International Archives of Photogrammetry and Remote Sensing 33 (3/W4), Annapolis, MD, USA. (2001). [21] M.A. Brovelli, M. Cannata and U.M. Longoni “Managing and processing LiDAR data within GRASS,” Proceedings of the Open source GIS – GRASS users conference Trento, Italy, Sep. (2002). [22] K. Kraus, and N. Pfeifer, “Adva nced DTM generation from LIDAR data,” International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 34 (Part 3/w4), 23 30, (2001). [23] M. Elmqvist, E. Jungert, F. Lantz, A. Perss on, and U. Sderman, “Terrain modeling and analysis using laser scanner data,” International Archives of Photogra mmetry, Remote Sensing and Spatial Information Sciences 34 (Part 3/W4), 211 218, (2001). [24] P. Axelsson, “DEM generation from lase r scanner data using adaptive TIN models,” International Archives of Photogrammetry and Remote Sensing 33 (B4/1), 110 117, Istanbul, Turkey, (2000). [25] G. Sithole and G. Vosselman, “Experimental comparison of filter algorithms for bare earth extraction from airborne laser scanning point clouds,” ISPRS Journal of Photogrammetry and Remote Sensing 59 85 101, (2004). [26] G. Sithole, “Segmentation and Classifi cation of Airborne Laser Scanner Data,” Dissertation, TU Delft ISBN 90 6132 292 8, Publications on Geodesy of the Netherlands Commission of Geodesy, 59 (2005). [27] J.L. Silvan-Cardenas and L. Wang, “A multiresolution approach for filtering LiDAR altimetry data,” ISPRS Journal of Photogrammetry and Remote Sensing 61 11 12, (2006).

PAGE 150

150 [28] W. Lu, J.J. Little, A. Sheffer, and H. Fu, “Deforestation: Extracting 3D Bare-Earth Surface from Airborne LiDAR data,” IEEE Canadian Conference on Computer and Robot Vision (2008). [29] X. Meng, L. Wang, J.L. Silvan-Cardenas, a nd N. Currit, “A multi-directional ground filtering algorithm for airborne LiDAR,” ISPRS Journal of Photogrammetry and Remote Sensing 64 117 124, (2009). [30] L. Y. Mimbela and L. A. Klein, “Summary of Vehicle Detection and Surveillance Techniques used in Intelligent Transportation Systems,” Federal Highway Administration’s (FHWA) Intelligent Transportation Systems Joint Program Office (2000). [31] JAMAR Technologies Inc., Catalog Number 7, Horsham, PA. [32] R.L. Gordon, R.A. Reiss, H. Haenel, E.R. Case, R.L. French, A. Mohaddes, and R. Wolcott, “ Traffic Control Systems Handbook ,” FHWA-SA-95-032, Federal Highway Administration, U.S. Department of Transportation, Washington, D.C., Feb. (1996). [33] H. Don, Finger on the Pulse Piezos on the Rise Traffic Technology International, June/July, (1999). [34] J.H. Kell and I.J. Fullerton, Traffic Detector Handbook Second Edition, U.S. Department of Transportation, Federal Highway Administration, Washington, D.C., (1990). [35] J.E. Lenz, FAA Demonstration of Alternate Technologies to Prevent Runway Incursions, Final Rpt 1 Honeywell Inc., Systems and Research Center, Minneapolis, MN, Apr. (1993). [36] H.R. Sampey, Vehicle Magnetic Imaging Nu-Metrics, Vanderbilt, PA, May (1994). [37] ASTM E1318-94, Standard Specification for Highway Weigh-in-Motion (WIM) with User Requirements and Test Method Annual Book of ASTM Standards, 04 (03), West Conshohocken, PA, (1994). [38] E.C. Chang, and H. Kunhuang, “Incident De tection Using Advanced Technologies,” Paper 930943, 72nd Annual Meeting, Trans portation Research Board Washington, D.C., (1993). [39] Nestor Corp. Data Sheets Providence, R.I., (1999). [40] L.A. Klein, Data Requirements and Sensor Technologies for ITS Norwood, MA, Artech House, (2001). [41] L.A. Klein and M.R. Kelley, Detection Technology for IVHS, Vol. I: Final Report FHWA-RD-95-100, Federal Highway Administratio n, U.S. Department of Transportation, Washington,D.C., Dec. (1996). [42] Y. Kumagai, T. Yamamoto, M. Deguchi, and S. Yamaoka, Ultrasonic Detector and New Type Sensors for Urban Traffic Control Systems Sumitomo Electric Industries, (1992). [43] M. Bertozzi, A. Broggi, and A. Fascioli, “Vis ion-based intelligent vehicles: State of the art and perspectives,” Robotics and Autonomous Systems 32 (2000) [44] Z. Sun, G. Bebis, and R. Miller, “On-Road Vehicle Detection: A Review,” IEEE Transactions on Pattern Analysis and Machine Intelligence 28 (5), 694 711, (2006).

PAGE 151

151 [45] M. Bertozzi, A. Broggi, M. Cellario, A. Fascioli, P. Lombardi, and M. Porta, “Artifical Vision in Road Vehicles,” Proc. IEEE 90 (7), 1258 1271, (2002). [46] F. Heimes and H. Nagel, “Towards Active Machine-Vision-Based Driver Assistance for Urban Areas,” Int’l J. Computer Vision 50 5 34, 2002. [47] U. Franke, D. M. Gavrila, A. Gern, S. Grzig R. Janssen, F. Paetzold and C. Whler, “From Door to Door Principles and Applications of Computer Vision for Driver Assistant Systems,” chapter 6 in Intelligent Vehicle Technologies eds. L. Vlacic and F. Harashima and M. Parent, Butterworth Heinemann, Oxford, (2001). [48] R. Bishop, Intelligent Vehicle Applications Worldwide IEEE Intelligent Systems, (2000). [49] M. Herbert, “Active and Passive Range Sensing for Robotics,” Proc. IEEE Int’l Conf Robotics and Automation 1 102 110, (2000). [50] S. Park, T. Kim, S. Kang, and K. Heon, “A Novel Signal Processing Technique for Vehicle Detection Radar,” IEEE MTT-S Int’l Microwave Symp. Digest 607 610, (2003). [51] C. Wang, C. Thorpe, and A. Suppe, “Ladar-Based Detection and Tracking of Moving Objects from a Ground Vehicle at High Speeds,” Proc. IEEE Intelligent Vehicles Symp. (2003). [52] J. Hancock, E. Hoffman, R. Sullivan, D. Ingimarson, D. Langer, and M. Hebert, “High-Performance Laser Ranger Scanner,” Proc. SPIE Int’l Conf. Intelligent Transportation Systems (1997). [53] R. Chellappa, G. Qian, and Q. Zheng, “Vehic le Detection and Tracking Using Acoustic and Video Sensors,” Proc. IEEE Int’l Conf. Acoustics, Speech, and Signal Processing 793 796, (2004). [54] Divide and conquer algorithm Wikipedia, the free encyclopedia, http://en.wikipedia.org/wiki/ Divide_and_conquer_algorithm (accessed 05.01.2010). [55] C.K. Toth, D.G. Brzezinska, and C. Merry “Supporting Traffic Flow Management with High-Definition Imagery,” Proceedings of Workshop on High Resolution Mapping from Space (2003). [56] C.K. Toth, D.G. Brzezinska, and S. Moafi poor, “Precise Vehicle Topology and Road Surface Modeling Derived from Airborne LiDAR Data,” The 60th Annual Meeting of ION 2004 Dayton, OH, June 7–9, (2004). [57] G. Vosselman, ISPRS Test On Extracting DEMs From Point Clouds: A Comparison Of Existing Automatic Filters ISPRS Commission III, http://www.itc.nl/isprswgIII-3/filtertest August 5th, (2003). [58] R. Shrestha, W. Carter, K.C. Slatton, and W. Dietrich, ‘Research-Quality’ Airborne Laser Swath Mapping: The Defining Factors National Center for Airborne Laser Mapping (NCALM), http://www.ncalm.ufl.edu/ Jun. 22, (2007). [59] Google earth, http://earth.google.com/ (accessed 04.15.2010) [60] M.A. Brovelli, M. Cannata, and U.M. Longo ni, “Managing and processing LiDAR data within GRASS,” Proceedings of the Open source GIS GRASS users conference Trento, Italy, Sep.

PAGE 152

152 (2002). [61] R.C. Gonzalez, and R.E. Woods, Digital Image Processing Prentice-Hall Publishers, New Jersey, 547 560 (2002). [62] R. J. Fowler and J. J. Little, “Automatic ex traction of irregular networ k digital terrain models,” Proceedings of the 6th annual conference on Computer graphics and interactive techniques 199–207, (1979). [63] R.C. Gonzalez and R.E. Woods, Digital Image Processing 2nd ed. Prentice Hall, 2002. [64] R.G. Congalton, “A review of assessing the accuracy of classifications of remotely sensed data,” Remote Sensing of Environment 37 35–46. (1991). [65] R.G. Congalton, and K. Green, Assessing the accuracy of remote ly sensed data: principles and practices Boca Raton, FL, USA: Lewis Publishers. 137 p. (1999). [66] F. Galton. Finger Prints Macmillan, London (1892). [67] J.R. Landis and G. G. Koch, “The measurem ent of observer agreement for categorical data,” Biometrics 33 159–174, (1977). [68] G. Sithole and G. Vosselman, Report: ISPRS Comparison of Filters ,. http://www.itc.nl/isprswgiii-3 /filtertest/report.htm (2003) (accessed 10.15.2009). [69] X. Meng, L. Wang, and N. Currit, “Morphology-based building detection from airborne LiDAR data,” Photogramm. Eng. Remote Sens. 75 427 442, (2009). [70] H. Arefi, J. Engels, M. Hahn, and H. Ma yer, “Automatic DTM generation from laser-scanning data in residential hilly area,” Proceedings of ISPRS Joint Workshop: Visualization and Exploration of Geospatial Data Stuttgart, Germany, June (2007). [71] A. Kobler, N. Pfeifer, P. Ogrinc, L. Todorovski, K. Oštir, and S. Deroski, “Repetitive interpolation: a robust algorithm for DTM generati on from aerial laser scanner data in forested terrain,” Remote Sens. Environ 108 9 23, (2007). [72] C. Wang, M. Menenti, M.P. Stoll, F. Alessandr a, B. Enrica, and M. Marco, “Separation of Ground and Low Vegetation Signatures in LiDAR Measurements of Salt-Marsh Environments,” IEEE Trans. Geosci. Remote Sens. 47 2014 2023, (2009). [73] J.L. Silvn-Crdenas and L. Wang, “A multiresolution approach for filtering LiDAR altimetry data,” ISPRS J. Photogramm. Remote Sens 61 11 22, (2006). [74] K. Zhang and D. Whitman, “Comparison of thr ee algorithms for filtering airborne LiDAR data,” Photogramm. Eng. Remote Sens 71 313 324, (2005). [75] Q. Zhou and X. Liu, “Error analysis on grid-based slope and aspect algorithm,” Photogramm. Eng. Remote Sens. 70 (8), 957 962, (2004). [76] M. Bartels and H. Wei, “Threshold-free object and ground point separation in LiDAR data,” Pattern Recognition Letters 31 (10), 1089 1099, July, (2010)

PAGE 153

153 [77] M. Bartels, H. Wei, and D.C. Mason, “DTM Generation from LIDAR Data using Skewness Balancing,” 18th International Conference on Pattern Recognition (ICPR) 1 Hong Kong, August 20 24, (2006). [78] Cluster analysis, http://en.wikipedia.org/wiki/Cluster_analysis (accessed 4.15.2010). [79] Object recognition, http://en.wikipedia.org/wiki/Object_recognition (accessed 4.15.2010). [80] Scale invariant feature transform, http://en.wikipedia.org/wiki/Scale-invariant_feature_transform (accessed 4.15.2010). [81] K-means clustering, http://en.wikipedia.org/wiki/K-means (accessed 4.15.2010). [82] Principal component analysis, http://en.wikipedia.org/wiki/Principal_components_analysis (accessed 4.15.2010). [83] R.O. Duda, P.E. Hart, and D.G. Stork, Pattern Classification 2nd ed. Wiley-Interscience, (2001). [84] A. Johnson, Spin-Images: A Representation for 3-D Surface Matching doctoral dissertation, The Robotics Institute, Carnegie Mellon Univ., (1997). [85] B. W. Silverman, Density Estimation for Statistics and Data Analysis Chapman and Hall, London, (1986). [86] S. Theodoridis, A Pikrakis, K. Koutroumbas, and D. Cavouras, Introduction to Pattern Recognition: A Matlab Approach Academic Press, UK, (2010). [87] C. Mallet, U. S¨orgel, and F. Bretar, “Analysis of full-waveform lidar data for classification of urban areas,” Photogrammetrie Fernerkundung GeoInformation (PFG) 5 337–349, (2008). [88] D. Gatziolis and H.E. Andersen, A Guide to LIDAR Data Acqu isition and Processing for the Forests Gen. Tech. Rep. PNWGTR768. Portland, (2008). [89] B. Jutzi and H. Gross, “Normalization of lidar intensity data based on range and surface incidence angle,” Laserscanning, International Archives of Photogrammetry and Remote Sensing 38 (Part 3/W8), 213 218, (2009). [90] C.K. Toth, A. Barsi, and T. Lovas, “Vehicle recognition from LiDAR data,” International Archives of Photogrammetry and Remote Sensing 34 (Part 3/W13), 162 166, (2003). [91] T. Lovas, C. K. Toth, and A. Barsi, “Model-based vehicle detection from LiDAR data,” 20th ISPRS Congress Commission 2, Istanbul, Turkey: ISPRS, (2004)

PAGE 154

154BIOGRAPHICAL SKETCH Li-Der Chang was born November 1971 in Taiwan and received the Bachelor of Science and Master of Science degree in electronic engineering from the Chung Cheng Institute of Technology, Taoyuan, Taiwan, R.O.C., in 1993 and 1999, respectively. In 1993-2006 (except two years while pursuing the Master of Science degree), he was a se nior research engineer in the telecommunication research and development de partment of the communication development office and the scientific research office in Taiwan. Since 2006, he has been working toward the Ph.D. degree in the Adaptive Signal Processi ng Laboratory of the El ectrical and Computer Engineering Department, University of Florida. His research in terests include remote sensing application, pattern recognition, al gorithm development, target detection, image processing, and adaptive signal processing.