GenTrack: Generalized Multi-Object Tracking with 3D Motion and Visual Appearance

다국어 초록 (Multilingual Abstract)

Title: GenTrack: Generalized Multi-Object Tracking with 3D Motion and Visual Appearance
Multi-object tracking (MOT) is a fundamental task in computer vision, with wide-ranging applications in areas such as autonomous driving, video surveillance, sports analysis, and industrial manufacturing. Although numerous models and algorithms have achieved promising results in object detection and tracking, several challenges remain significant obstacles, including crowded scenes, occlusions, camera motion, and high frame rates. Most existing approaches are developed and evaluated on standardized public datasets with low levels of distortion. However, real-world deployments, particularly in surveillance systems utilizing fisheye lens CCTV cameras with wide fields of view, often involve significant image distortion. This distortion is rarely addressed in traditional methods because it is largely absent from public datasets but common in practical surveillance scenarios.
In this study, we propose a more generalized tracking framework designed to effectively handle occlusions, crowded environments, and lens distortion, thereby improving tracking performance over existing methods. Our approach leverages motion cues to address issues related to crowded scenes and occlusion, while incorporating appearance cues to manage challenges such as high frame rates and rapid object movements. Additionally, the proposed method demonstrates strong generalization and achieves robust performance not only on standard public benchmarks but also on more challenging datasets characterized by high levels of distortion. This is achieved through the use of nonlinear transformations to localize objects in the 3D world, enabling more stable and accurate tracking.

목차 (Table of Contents)

• Summary i
• List of Tables iii
• List of Figures v
• 1. Introduction 1
• 2. Related Work 3
• 2.1 Tracking-by-detection 3
• 2.2 Tracking-by-regression 4
• 2.3 Tracking-by-segmentation 5
• 2.4 Tracking-by-end-to-end-models 5
• 3. GenTrack: Generalized Multi-Object Tracking with 3D Motion and Visual
• Appearance 7
• 3.1 Problem Formulation 7
• 3.2 GenTrack Overall Procedure 8
• 3.3 3D Object Localization 13
• 3.4 Adaptive Weighting Matching 17
• 4. Experiments 23
• 4.1 MOT Benchmark 23
• 4.2 Ablation Study 28
• 5. Conclusion 33
• References 34
• 국문초록 39
• 감사의 글 41