Luận án tiến sĩ: Bayesian Video Object Tracking - Dehong Ma
Luận án tiến sĩ về theo dõi đối tượng video sử dụng phương pháp Bayesian. Kết hợp mô hình với particle filtering xử lý occlusion, lighting changes và tracking multiple objects.
University of Wisconsin-Milwaukee
Kỹ thuật
Luan An
Luận án
Năm xuất bản
Số trang
120
Thời gian đọc
18 phút
Lượt xem
0
Lượt tải
0
Phí lưu trữ
40 Point
Mục lục chi tiết
Tóm tắt nội dung
I. Bayesian Video Object Tracking Overview
Video object tracking represents a critical challenge in computer vision. The technology enables automated surveillance, traffic monitoring, and video compression. Traditional methods struggle with lighting changes, occlusions, and non-rigid object morphology. This dissertation presents a Bayesian framework that addresses these fundamental challenges. The approach combines probabilistic tracking with particle filter implementation. It achieves minimum square error estimation of object states. The framework integrates motion detection with robust feature extraction. Shape and color features complement traditional kinetic parameters. This multi-modal approach enhances tracking reliability across diverse scenarios. Applications range from parking surveillance to unmanned aerial vehicle monitoring.
1.1. Core Research Problem
Video object tracking faces multiple technical obstacles. Over-segmentation and under-segmentation create tracking errors. Moving objects present different views across video frames. Non-rigid objects change shape during motion. Multiple object occlusions complicate state estimation. Lighting variations affect feature detection. Shadows and reflections introduce false positives. These challenges cause frequent object loss and high false alarm ratios. The research addresses these issues through Bayesian inference.
1.2. Proposed Solution Framework
The dissertation introduces a comprehensive Bayesian framework. A novel object model provides compact representation. The state vector combines shape, color, and kinetic features. Motion detection supplies observational data. State dynamic equations describe temporal evolution. Observation equations link measurements to hidden states. The framework employs particle filter for efficient implementation. This approach handles non-linear and non-Gaussian distributions effectively.
1.3. Key Contributions
The research delivers several innovative contributions. A unique state representation integrates multiple feature types. The model accommodates shape-changing objects naturally. Lighting robustness emerges from color feature processing. Multiple object tracking uses modified joint probability data association. Sequential likelihood testing enables automatic object initialization. The system reduces false alarms significantly. Track-loss frequency decreases through robust estimation.
II. Bayesian Framework for State Estimation
Bayesian inference provides optimal state estimation under uncertainty. The framework computes posterior distribution from prior probability and likelihood function. Prior probability encodes knowledge about object dynamics. Likelihood function measures observation consistency with hypothesized states. The posterior distribution represents complete knowledge given available data. Minimum mean square error estimation derives from the posterior mean. This approach handles uncertainty systematically. The Bayesian framework naturally incorporates multiple information sources. Motion models predict future states based on dynamics. Observation models connect measurements to underlying states. The recursive structure enables real-time processing. Each time step updates beliefs using new observations.
2.1. Prior Probability Modeling
Prior probability captures object motion characteristics. The motion model predicts state evolution over time. Kinetic features include position, velocity, and acceleration. Shape features describe object geometry. Color features represent appearance information. The prior distribution combines these elements coherently. State transition probabilities account for motion uncertainty. The model adapts to different object types and scenarios.
2.2. Likelihood Function Design
Likelihood function measures observation quality. Motion detection provides binary foreground masks. Color histograms characterize object appearance. Shape descriptors capture geometric properties. The likelihood combines multiple feature channels. Each channel contributes weighted evidence. Robust statistics handle outliers and noise. The function adapts to changing environmental conditions.
2.3. Posterior Distribution Computation
Posterior distribution combines prior and likelihood via Bayes rule. The computation updates beliefs recursively. Normalization ensures proper probability distribution. High-dimensional state spaces challenge direct computation. Particle filter provides approximate solution. The posterior mean gives optimal state estimate. Posterior variance quantifies estimation uncertainty. This information supports decision-making in tracking systems.
III. Particle Filter Implementation Methods
Particle filter enables efficient Bayesian inference for complex systems. The method represents posterior distribution through weighted samples. Each particle corresponds to a hypothesized state. Weights reflect relative likelihood of each hypothesis. The algorithm propagates particles according to motion model. Observations update particle weights through likelihood evaluation. Resampling eliminates low-weight particles and duplicates high-weight ones. This process maintains computational efficiency. Particle filter handles non-linear motion models naturally. Non-Gaussian distributions pose no theoretical difficulty. The approach scales to high-dimensional state spaces. Implementation requires careful tuning of particle count and resampling strategy.
3.1. Particle Propagation Strategy
Particle propagation predicts future states. The motion model generates proposal distribution. Each particle advances according to state dynamics. Random perturbations account for process noise. The proposal distribution should cover likely state regions. Importance sampling corrects for distribution mismatch. Effective particle placement improves estimation accuracy. Adaptive strategies adjust propagation based on tracking confidence.
3.2. Weight Update Mechanism
Weight update incorporates new observations. Likelihood function evaluates each particle. Weights multiply by observation probability. Normalization maintains sum-to-one constraint. High-weight particles represent probable states. Low-weight particles indicate unlikely hypotheses. Weight concentration indicates tracking confidence. Effective sample size monitors particle degeneracy.
3.3. Resampling Techniques
Resampling prevents particle degeneracy. The process eliminates particles with negligible weights. High-weight particles duplicate proportionally. Systematic resampling reduces sampling variance. Residual resampling preserves high-weight particles deterministically. Adaptive resampling triggers based on effective sample size. The procedure maintains particle diversity. Resampling frequency affects computational cost and estimation quality.
IV. Object Detection and Feature Extraction
Object detection identifies potential tracking targets in video frames. Motion detection separates foreground from background. Background subtraction provides initial object hypotheses. Morphological operations clean detection masks. Connected component analysis groups pixels into objects. Feature extraction characterizes detected objects. Shape features include area, perimeter, and moments. Color features use histogram representations. Texture features describe spatial patterns. Robust features ensure tracking stability across conditions. Feature selection balances discriminative power and computational cost. Multi-modal features improve tracking robustness. The detection and extraction pipeline feeds the Bayesian tracker.
4.1. Motion Detection Algorithms
Motion detection identifies changing image regions. Background subtraction compares frames to reference model. Adaptive background models handle gradual changes. Gaussian mixture models represent multi-modal backgrounds. Frame differencing detects rapid motion. Optical flow estimates pixel displacement. Each method has specific advantages. Background subtraction works well for static cameras. The dissertation employs adaptive background modeling for robust detection.
4.2. Shape Feature Computation
Shape features describe object geometry. Area measures object size. Perimeter indicates boundary length. Aspect ratio characterizes elongation. Moments capture shape distribution. Central moments provide translation invariance. Normalized moments ensure scale invariance. Shape features track non-rigid object deformation. The compact representation reduces computational burden.
4.3. Color Feature Representation
Color features characterize object appearance. Histograms summarize color distribution. RGB color space provides direct representation. HSV space separates intensity from chromaticity. Histogram binning balances resolution and robustness. Normalization handles illumination changes. Color features distinguish similar-shaped objects. The representation remains stable under moderate lighting variations.
V. Multiple Object Tracking with JPDA
Multiple object tracking requires data association solutions. The problem assigns observations to tracked objects. Joint probability data association computes assignment probabilities. The algorithm considers all feasible associations. Each hypothesis receives probability weight. Weights depend on observation likelihood and prior track information. The method handles measurement uncertainty systematically. JPDA avoids hard assignment decisions. Probabilistic association reduces tracking errors. The dissertation proposes JPDA variation for video tracking. Modifications address computational efficiency. The adapted algorithm scales to moderate object counts. Implementation combines Bayesian state estimation with probabilistic association.
5.1. Data Association Problem
Data association matches observations to tracks. Multiple objects create ambiguous measurements. Occlusions generate missing observations. Clutter produces spurious detections. The combinatorial problem grows exponentially. Optimal solution requires exhaustive enumeration. Practical algorithms use approximations. Joint probability data association provides probabilistic framework. The method marginalizes over association uncertainties.
5.2. JPDA Algorithm Adaptation
The dissertation modifies standard JPDA for video tracking. Gating reduces computational complexity. Distance thresholds eliminate unlikely associations. Track-to-observation likelihood uses multiple features. Shape and color complement position information. The algorithm computes joint association probabilities. Marginal probabilities weight individual associations. State updates incorporate weighted observations. The adaptation maintains real-time performance.
5.3. Occlusion Handling Strategy
Occlusions challenge multiple object tracking. Partial occlusions reduce observation quality. Complete occlusions eliminate measurements temporarily. The system maintains track predictions during occlusions. Motion model propagates state forward. Confidence decreases without observation updates. Track deletion delays prevent premature termination. Re-detection mechanisms recover occluded objects. The strategy reduces track fragmentation significantly.
VI. Automatic Initialization and Deletion
Fully automatic tracking requires initialization and deletion mechanisms. Sequential likelihood testing provides principled approach. The method accumulates evidence for object presence. Likelihood ratios compare object and no-object hypotheses. Sequential testing allows early decision making. Initialization thresholds balance sensitivity and specificity. New objects receive tentative tracks. Confirmation requires sustained detection. Deletion removes lost or exited objects. Track confidence monitoring triggers deletion decisions. The technique reduces false alarm ratio substantially. Track-loss frequency decreases through careful threshold tuning. Automatic management enables long-term surveillance applications.
6.1. Sequential Likelihood Test Theory
Sequential likelihood test evaluates competing hypotheses. The method computes cumulative log-likelihood ratio. Upper threshold confirms object presence. Lower threshold rejects object hypothesis. Intermediate values continue testing. The approach minimizes expected sample size. Decision boundaries control error rates. False positive and false negative probabilities balance. The test adapts to observation quality. Strong evidence enables rapid decisions.
6.2. Object Initialization Process
Initialization creates new tracks from detections. Unassociated observations generate track candidates. Sequential testing evaluates candidate validity. Likelihood accumulates over multiple frames. Consistent detections increase confidence. Shape and color features verify object identity. Confirmed tracks enter active tracking. The process filters transient false detections. Initialization delay trades latency for accuracy.
6.3. Track Deletion Criteria
Track deletion removes invalid or terminated tracks. Observation absence decreases track confidence. Sequential test monitors cumulative evidence. Deletion threshold triggers track removal. The system distinguishes occlusions from exits. Spatial context provides additional information. Edge proximity suggests object departure. Deletion prevents resource waste on lost objects. Careful tuning maintains track continuity during temporary occlusions.
Tải xuống file đầy đủ để xem toàn bộ nội dung
Tải đầy đủ (120 trang)Trích đoạn nội dung luận án
Tải xuống để đọc toàn bộBAYESIAN VIDEO OBJECT TRACKING By Dehong Ma A Dissertation Submitted in Partial Fulfilment of the Requirements for the Degree of Doctor of Philosophy In Engineering at The University of Wisconsin-Milwaukee December 2006 UMI Number: 3244154 INFORMATION TO USERS The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleed-through, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion.
® UMI UMI Microform 3244154 Copyright 2007 by ProQuest Information and Learning Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. ProQuest Information and Learning Company 300 North Zeeb Road P.
Box 1346 Ann Arbor, MI 48106-1346 BAYESIAN VIDEO OBJECT TRACKING By Dehong Ma A Dissertation Submitted in Partial Fulfilment of the Requirements for the Degree of Doctor of Philosophy In Engineering at The University of Wisconsin-Milwaukee December 2006 ose /2 16 ener Major-Prefessor—— mĩ ate i) Graduate School Approval. Date 11 BAYESIAN VIDEO OBJECT TRACKING By Dehong Ma The University of Wisconsin-Milwaukee, 2006 Under the Supervision of Professor Jun Zhang This is a study of tracking moving objects robustly and efficiently from a video stream. Video object tracking aids higher level image analysis by providing richer image information more: efficiently. It can find many applications in smart video surveillance, video conferencing, human computer interface, traffic measurement, image stabilization, video compression, etc.
However, it is still an open problem due to the difficulties from over/under segmentations, the different views of a moving obj ect, the morphology of non- rigid objects, the occlusions of multiple moving objects, lighting changes, shadows, reflections, etc. These difficulties often result in frequent object loss, high false alarm ratio, and other problems. In this thesis we describe a Bayesian framework, combining with a novel model and an efficient particle filtering implementation, to combat those difficulties. The Bayesian 11 approach is optimal in that it gives the minimum square error (MSE) estimation of the object’s states being tracked.
Because the key issue of this approach is modelling, a unique model was proposed and studied in this research. It consists of a compact representation of a moving object, a unique state vector formed from robust shape and colour features in addition to the often-used kinetic features, observations from motion detection, and their relationships described by the state dynamic and observation equations. This model has the advantages of being able to track non-rigid or shape- changing objects and being robust with lighting variations. In order to track multiple objects simultaneously, we studied the typical multiple hypothesis tracking algorithms and proposed a variation of the joint probability data association (JPDA) algorithm to solve the data association problem in video object tracking.
For-fully automatic tracking, a new technique based on sequential likelihood test is combined into the system for the object initialisation and deletion. This technique can greatly reduce the false-alarm ratio and track-loss frequency. The robustness and efficacy of the system are demonstrated by tracking various objects in a variety of applications, such as video based parking-plot surveillance, and video tracking from Unmanned Aerial Vehicles (UAVs). a _Major Professor _/2zl4 lo Date 1V Acknowledgments My foremost gratitude goes to Professor Jun Zhang, my advisor, for his academic and financial support throughout my Ph.
This thesis could not have been completed without his guidance, insightful instructions, and constructive criticism. He is a model of a teacher and a supervisor besides a perfect academic advisor. Over the years, I have learned from him not only effective research approaches and good study habits, but also academic writing and an optimistic, positive attitude. My appreciation also goes to Professor Gilbert G.
Walter, one of my committee members, for his help on my study and research, and Professor Chiu Tai Law, also a member of my-program committee, for all his help in my Ph. study in UWM. ! would like to express my thanks to Professor D. Hosseini, and Professor IG.
Lauko for all their help in my doctoral study and being members of my program committee. I would like to give my thanks to my colleagues and dear friends Jianbo Gao, Xiao Zhang, Weisong Liu and Yirong Wu, Xin Sheng, Wen Hu, Chuan Zhou, Sun Tong, Jieping Xu and all other friends for your help and support in the period of my doctorate study. Weisong deserves my special thanks for his help in the coding of the thesis project. Finally, I must give my deepest gratitude and love to my wife, Ruiyun Wang, who supported me with all her efforts during the hard times in my doctoral study.
Without her endless love and assistance to the family, it would be impossible for me to finish this thesis. Hence, I would like to dedicate it to her, and my dear daughter Jenny, my parents Yiju Ma and Nianxiu Li, and all my family for ever-lasting encouragement, inspiration, and love! Milwaukee, December 2006 Dehong Ma vi Table of Contents Acknowledgements V 1.2 Definition of Video Object TTAaCKINE. HH ee sees kh, 3 1.3 Difficulties of Video Object Tracking and Objectives of Research.4 Outline b09:(-~*tta|iIÍIAIỌỤỌIAIiẢ. Previous Work on Video Object Tracking 8 2.
¬ eee ne ene EERE EOE READE EES HEEEEE EO DES ED EDP EE DESO SEA ee ESOS EEE EES EEOE EES 8 2. nh nà erent nh nà ki nà hệt 9 2. cu HH nh nh nh kh re 10 2.2 Motion Def€CtIOn. cQQn n nnHn nH nh nh nh hy, 12 2.
Representation, Modelling, and Methods -of Filtering.4 Data Association Techmiques. HQ nh nh nhe 18 2.5 Objectives of this Research.c con ng ng HH kh hs. Object Detection Based on Non-linear Prediction 23 EM si Ji gdaiiaaẳaiiiiiiiiiiiiii.2 Gaussian Mixture Model and Its Optimal Prediction.1 Multivariate Gaussian ï nn "¬""— 24 3.2 Optimal Prediction of Multivariate Gaussian Mixture. Small Object Detection Based on Non-linear Optimal Prediction.4 Experimental Results and Comparison with AR model.
Bayesian Single Video Object Tracking 4. ni eee ene teen nh nh nh nà nh bà 4. A Robust Model for Video Object Tracking.1 System State and System Dynamic Mode.2 Observation and Observation Model. ng nà sees eaeaeneeeeeeenes 4.
vet eceeeee pe eneeee eee teeeseeeaene tenes 4. Experimental Results and Conclision. Bayesian Multiple Video Object Tracking 5. ccc cece eee neee cere eneee teen eneeeeeeneeeeneneeteeeeaenees 5.2 Bayesian Filter for Multiple Object Tracking .1 The three steps of Multiple Hypothesis Tracking.3 Hypothesis Likelihood Evaluation and Models.4 Bayesian Multiple Hypothesis Tracking Formulations 5.
Sequential Likelihood Ratio Test for Object Maintenance."— eden serene e neces ene eea ene na enes 5.cc cc cn ee eee eeeneeeseeeaenaeseenens Vill 5.4 Data Association for Multiple Video Object Tracking.1 Multiple Hypothesis Tracking (MHT).2 Joint Probabilistic Data Association Filter (TPDAF). Probabilistic Multiple Hypothesis Tracking (PMHT).5 Experiment Results and Conclusion. cà nền nha 85 5.1 Simulation Experiments and Comparison of MHT, JPDA, and PMHT. Experiments on Multiple Video Object Tracking Using MHT, JPDA, and 6.
Summary 93 1X List of Tables Table 3. Implementation/Control pararnef€TS. MSEEs from MHT, JPDAF, and PMHT on simulation. 88 List of Figures Figure 1.
The states (x) and observations (z) in a tracking system. A typical video tracking SySf€T.- Ác HH ng ng m 9 Figure 2. The grouped observations are from objects. Isolated are from false alarms.
con HH ng HH ee ene ence eee km vàn 20 Figure 3.Right: one-dimension vector. An Object in Synthetic Background. 00sec cee e eee e eee rere eee ee ee 29 Figure 3. An Object in a Brodatz Texture Background.
A Real World Ímage.--- ng nh hy 33 Figure 3. Visualizing the Optimal Nonlinear and Linear Predictors - A Simple ExampÏÌ©. eee eee eee cee eee sen cee tee een eee keo se «uc Figure 4. Illustration of the state of a moving obJecf.
A typical moving region detection example by our technique. Here the images were taken from a MOVING CAMETA.ccceceeee ee eee eee ee eees 51 Figure 4. Illustration of the observation data. Tracking a single person using a PTZ camera on pre-recorded video clip.
Red ellipse: our result; Green rectangle: template matching; Blue rectangle: histogram matching (mean-shIfẨ).-- cà taeeeeaeens 61 Figure 4. Tracking a single person in real-time using a PTZ camera. Red ellipse: tracking result; Blue ellipse: partIcles.1: Illustration of hypothesis generating. Simulation experiments on MHT, JPDAF, and PMHT.
Real video experiments using MHÍT. sành he 90 XI Figure 5. Real video experiments using JPDAF. eee eee ee ee eed Figure 5.
Real video experiments using PMHT ee ee ry xii Chapter 1 Introduction 1.1 Background “Vision is knowing what is where by looking,” D. Marr said in his book Vision [1]. The ultimate goal of computer vision, or machine vision, is to make a computer/machine be able to “see” the world. Computer vision is one of the most important aspects of machine intelligence.
The research on computer vision has evolved from laboratories to the real world by finding many applications in a variety of areas, such as in medical systems [2]-[6], factory automation [7]-[10], remote sensing [11]-[16], bio-identification in security and law enforcement [17]-[25], smart vehicles [26]-[30], etc. Almost all of the modern intelligent machines resort to machine vision to acquire outside information as a smart input device. In order to provide richer image information more efficiently, video object tracking, especially active object tracking by controlling sensor parameters [31]-[33], is a critical process in a computer vision system. A typical vision perception process begins with motion detection, also called focus invoke, then it has a fixation or tracking process to b2 acquire more information about the moving object until an assessment or a further action has been made, which ends a perception circle.
In addition to acquiring better image information of the object of interest, video object tracking itself finds many direct applications. In smart surveillance or virtual guard systems it is used for event detection and compression [35]-[38], and in teleconferencing for auto-recording [39]-[40]. It is also used for traffic analysis [41]-[42], human-computer interface [43]-[45], image stabilization [46]-[48], automatic target tracking in UAV [49]- [51], and image registration in medical imaging [6] [52]. In recent years the need for smart surveillance for homeland security, crime prevention and verification has strengthened.
Usually a smart surveillance system should be able to detect potential criminal activities in a public area and to obtain close-up video recordings and give instant warnings when suspicious behaviours have been detected. Currently these tasks are mostly done with human participation, which is labour-intensive and stressful, and also creates a variability between different operators (due to experience and work ethic). Different times of the day also can cause inconsistent results. All these applications call for reliable, robust, efficient and fully automatic video object tracking techniques, which are still an open research area due to a variety of difficulties remaining unresolved.2 Definition of Video Object Tracking Computationally, the tracking process is to estimate the state (position, velocity, etc.) of a moving object from its observations or measurements over time [53].
The video object tracking is to estimate a moving object’s state through its video observations. The observations are usually “noisy,” which can be caused by observation inaccuracy or measurement errors. Also, there is uncertainty about the state itself at any given time, so the state can be viewed as a random process over time. Hence, tracking is essentially a statistical estimation process of the state of a random process [54].
A tracking system mainly consists of two elements: system state and system observation. The system state refers to the state of the object of being tracked, which usually consists of a set of state variables. They are also called the system state-vector. Some often-used state variables are an object’s kinetic parameters.
Nội dung được bảo vệ bản quyền — Tải xuống đầy đủ
Từ khóa và chủ đề nghiên cứu
Câu hỏi thường gặp
Luận án "Luận án tiến sĩ Bayesian Video Object Tracking" nghiên cứu về vấn đề gì?
Luận án tiến sĩ về theo dõi đối tượng video sử dụng phương pháp Bayesian. Kết hợp mô hình với particle filtering xử lý occlusion, lighting changes và tracking multiple objects.
Luận án "Luận án tiến sĩ Bayesian Video Object Tracking" được bảo vệ tại trường nào?
Luận án này được bảo vệ tại University of Wisconsin-Milwaukee. Năm bảo vệ: 2006.
Luận án "Luận án tiến sĩ Bayesian Video Object Tracking" thuộc chuyên ngành gì?
Luận án "Luận án tiến sĩ Bayesian Video Object Tracking" thuộc chuyên ngành Kỹ thuật. Danh mục: Khoa Học Máy Tính.
Luận án "Luận án tiến sĩ Bayesian Video Object Tracking" có bao nhiêu trang?
Luận án "Luận án tiến sĩ Bayesian Video Object Tracking" có 120 trang. Bạn có thể xem trước một phần tài liệu ngay trên trang web trước khi tải về.
Cách tải luận án "Luận án tiến sĩ Bayesian Video Object Tracking" về máy như thế nào?
Để tải luận án về máy, bạn nhấn nút "Tải xuống ngay" trên trang này, sau đó hoàn tất thanh toán phí lưu trữ. File sẽ được tải xuống ngay sau khi thanh toán thành công. Hỗ trợ qua Zalo: 0559 297 239.