ebook img

Person Re-Identification And Intelligent Crowdsourcing With Applications In Public Safety PDF

187 Pages·2017·15.48 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Person Re-Identification And Intelligent Crowdsourcing With Applications In Public Safety

PERSON RE-IDENTIFICATION AND INTELLIGENT CROWDSOURCING WITH APPLICATIONS IN PUBLIC SAFETY A Dissertation Submitted to the Faculty of Purdue University by Khalid Tahboub In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy August 2017 Purdue University West Lafayette, Indiana ii THE PURDUE UNIVERSITY GRADUATE SCHOOL STATEMENT OF DISSERTATION APPROVAL Dr. Edward J. Delp, Chair School of Electrical and Computer Engineering Dr. Amy R. Reibman School of Electrical and Computer Engineering Dr. Mary L. Comer School of Electrical and Computer Engineering Dr. Zygmunt Pizlo Department of Psychological Sciences Approved by: Dr. Venkataramanan Balakrishnan Head of the School Graduate Program iii To my parents, Khulusi & Majeda, with my deepest gratitude iv ACKNOWLEDGMENTS First and foremost, I would like to thank my doctoral advisor Professor Edward J. Delp. This thesis would not have been possible without his encouragement, support, guidance and criticism. I am grateful to him for the opportunity and privilege of becoming a member of the Video and Image Processing Laboratory (VIPER). As a Ph.D. student, Professor Delp has been my role mode. His technical depth and breadth, ability to inspire, and encouragement to overcome obstacles have been truly invaluable. During our numerous conference trips, the many drives to Chicago and Cleveland, I have also enjoyed our nonacademic-related conversations from which I learned a lot. I would like to thank Professor Amy R. Reibman for her guidance and support. I am very grateful for the opportunity to work with a world class researcher. Her educational approach, attention to details, kindness and analytical skills have made working with her an enjoyable learning experience. I would like to thank Professor Mary Comer for her encouragement throughout my graduate studies. Her support and guidance have been invaluable. I would also like to thank Professor Zygmunt Pizlo for his insightful feedback and suggestions. I would like to thank Dr. Neeraj Gadgil for being a great teammate for the crowdsourcing and the content based video retrieval projects. I am thankful for his friendship and support during my Ph.D. studies. I am also thankful to all of VIPER lab members for their encouragement, support and friendships. In particularly, I would like to thank Blanca Delgado and Dahjung Chung for their help in the vBOLO project and for being great teammates. I also want to thank Javier Ribera for his work on the crowdsourcing project and for being a great colleague. And thank you to the rest of my former and current colleagues: Di Chen, Qingchaung (Cici) Chen, Yuhao Chen, Jeehyun Choe, David Gu¨era Cobo, Shaobo Fang, Chichen Fu, Dr. Ye v He, David Ho, Joonsoo Kim, Soonam Lee, He Li, Chang Liu, Daniel Mas, Dr. Albert Parra Pozo, Ruiting Shao, Yu Wang, Dr. Chang (Joy) Xu, Sri Kalyan Yarlagadda, Jiaju Yue, Dr. Bin Zhao. I would also like to thank our lab visitors Thitiporn (Bee) Pramoun and Kharittha (Poy) Thongkur. I would like to thank the endowment of the Charles William Harrison Distinguished Professorship at Purdue University for partially supporting the work on the content based video retrieval project. The Chicago LTE test results discussed in the introduction chapter were obtained in cooperation with the Chicago Police Department, Motorola Solutions and the U.S. Department of Homeland Security. We gratefully acknowledge their cooperation. The images shown in the person re-identification chapter were obtained in co- operation with the Greater Cleveland Regional Transit Authority. We gratefully acknowledge their cooperation. This work was partially supported by the Cisco University Research Program Fund CG-#594368 through the Silicon Valley Community Foundation. This dissertation was partially supported by the Visual Analytics for Command, Control, and Interoperability Environments (VACCINE) center under award number 2009-ST-061-CI000. VACCINE is a U.S. Department of Homeland Security’s center of excellence at Purdue University. I am grateful for their support of the crowdsourcing project, vBOLO project and the Chicago LTE study. vi TABLE OF CONTENTS Page LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi 1 INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.1 Video Surveillance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Video Analytics in a Mobile or Networked Environment . . . . . . . . . 2 1.3 Person Re-Identification . . . . . . . . . . . . . . . . . . . . . . . . . . 7 1.4 Crowdsourcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1.5 Contributions Of The Thesis . . . . . . . . . . . . . . . . . . . . . . . . 10 1.6 Publications Resulting From Our Work . . . . . . . . . . . . . . . . . . 14 2 QUALITY-ADAPTIVE DEEP LEARNING FOR PEDESTRIAN DETECTION . . . . . . . . . . . . . . . . . . . . . . 17 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.2 Overview of Previous Work . . . . . . . . . . . . . . . . . . . . . . . . 18 2.3 Pedestrian Detection Using Compressed Video Sequences . . . . . . . . 21 2.4 Proposed Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.4.1 Training RPN Using Compressed Video Sequences . . . . . . . . 23 2.4.2 Proposed Two-Stage CNN . . . . . . . . . . . . . . . . . . . . . 25 2.4.3 Quality Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.5 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . . . 27 3 ACCURACY PREDICTION FOR PEDESTRIAN DETECTION . . . . . . . . . . . . . . . . . . . . . . . 31 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 vii Page 3.2 Pedestrian Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.3 Proposed Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 3.3.1 Texture Descriptors . . . . . . . . . . . . . . . . . . . . . . . . . 36 3.3.2 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 4 PERSON RE-IDENTIFICATION . . . . . . . . . . . . . . . . . . . . . . . . 45 4.1 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 4.1.1 Ensemble Of Local Features . . . . . . . . . . . . . . . . . . . . 49 4.1.2 Local Maximal Occurrence Feature . . . . . . . . . . . . . . . . 50 4.1.3 Hierarchical Gaussian Descriptor . . . . . . . . . . . . . . . . . 51 4.1.4 Keep It Simple and Straightforward Metric Learning . . . . . . 52 4.1.5 Cross-View Quadratic Discriminant Analysis . . . . . . . . . . . 53 4.2 Data Collection and Public Datasets . . . . . . . . . . . . . . . . . . . 53 4.3 Proposed Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 4.3.1 An Ensemble Of Localized Patches . . . . . . . . . . . . . . . . 62 4.3.2 Person Re-Identification Using A Patch-Based Appearance Model . . . . . . . . . . . . . . . . 66 4.3.3 Dual Linear Regression-Based Classification . . . . . . . . . . . 72 4.3.4 Metric Learning-Based Deformable Graph Matching . . . . . . . 74 4.4 Experimental Results and Discussion . . . . . . . . . . . . . . . . . . . 78 5 INTELLIGENT CROWDSOURCING . . . . . . . . . . . . . . . . . . . . . 88 5.1 Overview of Previous Work . . . . . . . . . . . . . . . . . . . . . . . . 88 5.2 System Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90 5.3 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 5.3.1 Administrator Portal . . . . . . . . . . . . . . . . . . . . . . . . 94 5.4 Crowdsourcing for Forensic Analysis of Surveillance Video . . . . . . . 98 5.4.1 Web-Based Annotation Platform with Fast Object Tracking . 100 5.4.2 Human Detection and Re-Identification . . . . . . . . . . . . . 102 viii Page 5.5 Crowdsourcing for Real-time Alerting . . . . . . . . . . . . . . . . . . 104 5.5.1 Crowd Flow Estimation . . . . . . . . . . . . . . . . . . . . . 104 5.5.2 Online Active Learning from Crowds . . . . . . . . . . . . . . 106 5.6 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 5.6.1 System Performance . . . . . . . . . . . . . . . . . . . . . . . 110 5.6.2 Trained Vs. Untrained Crowds . . . . . . . . . . . . . . . . . 111 5.6.3 Online Active Learning From Crowds . . . . . . . . . . . . . . 114 5.6.4 Crowdsourcing for Forensic Analysis of Surveillance Video . . 117 6 COMPRESSED DOMAIN VIDEO SIGNATURES . . . . . . . . . . . . . 121 6.1 Overview of Previous Work . . . . . . . . . . . . . . . . . . . . . . . 123 6.2 Proposed Video Signature . . . . . . . . . . . . . . . . . . . . . . . . 128 6.2.1 Motion Vectors in HEVC . . . . . . . . . . . . . . . . . . . . . 128 6.2.2 Projections on Random Matrices . . . . . . . . . . . . . . . . 129 6.2.3 Signature Generation . . . . . . . . . . . . . . . . . . . . . . . 131 6.2.4 Signature Matching . . . . . . . . . . . . . . . . . . . . . . . . 134 6.2.5 Video Retrieval From Mobile Devices: How Much Content Is Enough? . . . . . . . . . . . . . . . . . 137 6.3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 6.3.1 Signatures Matching . . . . . . . . . . . . . . . . . . . . . . . 139 6.3.2 Content Quantification . . . . . . . . . . . . . . . . . . . . . . 141 6.3.3 Video Retrieval Performance . . . . . . . . . . . . . . . . . . . 144 6.4 Conclusion and Future Work . . . . . . . . . . . . . . . . . . . . . . . 144 7 SUMMARY AND FUTURE WORK . . . . . . . . . . . . . . . . . . . . . 146 7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148 7.3 Publications Resulting From The Thesis . . . . . . . . . . . . . . . . 150 REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 ix LIST OF TABLES Table Page 3.1 Mean absolute errors for miss rate prediction . . . . . . . . . . . . . . . . . 39 4.1 Spatial resolutions and frame rates used for camera locations in GCRTA . 54 4.2 Summary of the videos used for ground truthing effort - Triskett Bridge . . 57 4.3 Summary of the videos used for ground truthing effort - West177 Tunnel . 57 4.4 Summary of testing data - Triskett Bridge . . . . . . . . . . . . . . . . . . 59 4.5 Summary of testing data - West177 Tunnel . . . . . . . . . . . . . . . . . . 59 4.6 Matching Accuracy - Triskett Bridge . . . . . . . . . . . . . . . . . . . . . 79 4.7 Matching Accuracy - West117 Tunnel . . . . . . . . . . . . . . . . . . . . . 79 5.1 Overall performance - crowdsourincg single video surveillance tasks . . . 111 5.2 Training performance - trained vs. untrained crowds . . . . . . . . . . . 112 5.3 Tasks summary - trained vs. untrained crowds . . . . . . . . . . . . . . . 113 5.4 Results - trained vs. untrained crowds . . . . . . . . . . . . . . . . . . . 113 5.5 Crowd flow estimation error rate with no crowdsourcing under video qual- ity degradation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 5.6 Crowd flow estimation error rates using the second type of uncertainty characterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115 5.7 Two approaches to aggregate crowdsourcing output and its impact on crowd flow estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116 5.8 Crowdsourcing for forensic analysis of surveillance video - investigation details . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 5.9 Crowdsourcing for forensic analysis of surveillance video - human detection results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118 5.10 Crowdsourcing for forensic analysis of surveillance video - human re-identification - automatic method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 5.11 Crowdsourcing for forensic analysis of surveillance video - human re-identification - using crowd . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120 x LIST OF FIGURES Figure Page 1.1 CPD district 7 borders . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.2 Police laptop connected to LTE using a USB dongle . . . . . . . . . . . . . 4 1.3 Locations used for subjective measurements within CPD district 7 . . . . . 5 1.4 Clarity, fluidity and video data rate along the cell edges - RTVI . . . . . . 6 1.5 Subject intrinsic variations and person re-identification challenges . . . . . 7 1.6 Sample sequences from two different cameras . . . . . . . . . . . . . . . . . 8 2.1 Miss rate versus FPPI for the RPN pedestrian detector . . . . . . . . . . . 22 2.2 Video compression and RPN performance . . . . . . . . . . . . . . . . . . 23 2.3 RPN detectors trained using QP {25,35,47,55} . . . . . . . . . . . . . . . . 24 2.4 RPN detectors trained using QP {20,35,47} . . . . . . . . . . . . . . . . . 25 2.5 Proposed Two-Stage Neural Network Block Diagram . . . . . . . . . . . . 26 2.6 Class assignments for training the Inception-v3 quality estimation network 27 2.7 Classification error rate as a function of the temporal window size k . . . . 28 2.8 Bank of RPN detectors performance . . . . . . . . . . . . . . . . . . . . . 30 3.1 ACF and LDCF miss rate versus QP . . . . . . . . . . . . . . . . . . . . . 35 3.2 Block diagram of accuracy prediction for pedestrian detection . . . . . . . 36 3.3 Prediction results for ACF and LDCF pedestrian detectors . . . . . . . . . 40 3.4 Prediction results for HOG and DPM pedestrian detectors . . . . . . . . . 41 3.5 Mean absolute error as a function of k . . . . . . . . . . . . . . . . . . . . 43 3.6 Mean absolute error for various number of responses . . . . . . . . . . . . 43 4.1 Subject intrinsic variations and person re-identification challenges . . . . . 45 4.2 Filters used to model texture . . . . . . . . . . . . . . . . . . . . . . . . . 50 4.3 GCRTA system map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54 4.4 Camera views at Brookpark, Puritas and Triskett sites . . . . . . . . . . . 55

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.