Unified Cross-Camera Learning for Person Re-Identification in Intelligent Surveillance Systems

다국어 초록 (Multilingual Abstract)

Modern multi-camera systems in public safety, retail analytics, and forensic search rely on person re-identification (Re-ID) to match individuals across non-overlapping views. In practice, Re-ID must cope with cross-camera domain shifts (viewpoint, illumination), foreground variability (pose, occlusion), and constraints such as limited labels and privacy budgets. This thesis studies cross-camera Re-ID under a unified setting and evaluation protocol spanning three regimes-fully unsupervised learning, unsupervised domain adaptation (UDA), and supervised learning-with consistent query–gallery retrieval and cross-camera CMC/mAP reporting.
We analyze when each regime is appropriate in deployment: fully unsupervised learning eliminates manual labeling on the target camera network but must construct reliable pseudo structures under severe camera bias; UDA transfers knowledge from a labeled source domain to an unlabeled target domain, reducing annotation cost while still suffering from domain shift and clustering noise; supervised learning can achieve the highest accuracy when rich labels are available but incurs substantial annotation overhead. Unifying these regimes clarifies the trade-offs among accuracy, labeling cost, robustness to domain shift, and privacy, and guides the design of scalable systems.
Accordingly, this thesis pursues three complementary aims: (Aim 1) enhance fully unsupervised Re-ID by learning robust cross-camera structures without iden tity labels; (Aim 2) improve UDA Re-ID by injecting mid-level semantic cues into clustering-based adaptation; and (Aim 3) strengthen supervised Re-ID by efficiently exploiting textual side information to emphasize identity-relevant visual cues.
Aim 1 is presented in Chapter 3 as the Optimal Proxy Mining Contrastive Network (OPMCN). OPMCN improves pseudo structures by Hardest Camera Proxy Mining to encourage camera-invariant representations, False Negative Proxies Mining to suppress erroneous negatives, and a PKGC block that couples global and local cues. On multiple large-scale datasets, OPMCN boosts unsupervised performance and robustness to camera bias.
Aim 2 is presented in Chapter 4 as an Auxiliary Attribute Clustering (AAC) framework for UDA Re-ID. A parsing-based attribute generator first produces image-level attribute labels, which are refined by an Attribute Label Refining strategy that removes detrimental categories, truncates over-confident scores, and rectifies incon-sistent labels. The refined attribute labels provide semantic similarity complementary to visual similarity, and AAC fuses both cues in a plug-and-play clustering module, yielding more reliable pseudo labels and competitive UDA performance.
Aim 3 is presented in Chapter 5 as the Text-Guided Fusion Transformer (TGFT) for the supervised regime. TGFT leverages a pretrained CLIP text encoder and a lightweight Gated Cross-Attention Fusion module that injects attribute-level se-mantics into the visual encoder, sharpening identity focus with minimal parameter overhead. Experiments on standard Re-ID benchmarks show consistent gains over strong CNN/ViT baselines.
Overall, these studies offer a unified view across unsupervised, UDA, and su-pervised regimes: high-quality proxies and false-negative control are crucial for label-free Re-ID; mid-level semantic attributes significantly enhance clustering-based domain adaptation; and text-guided fusion provides an efficient way to exploit human-understandable semantics in supervised training.

목차 (Table of Contents)

• Acknowledgements iii
• Abstract iv
• 1 Introduction 1
• 1.1 Motivation and Background 1
• 1.2 Problem Description and Objective 2
• 1.3 Contributions 3
• 1.4 Disposition 4
• 2 Literature Review 5
• 2.1 Datasets and Evaluation Protocols 5
• 2.2 Backbone Architectures and Training Recipes 6
• 2.3 Supervised Person Re-ID 6
• 2.4 Fully Unsupervised Person Re-ID 7
• 2.5 Unsupervised Domain Adaptation for Re-ID 8
• 2.6 Summary 8
• 3 Optimal Proxy Mining Contrastive Network 9
• 3.1 Introduction 9
• 3.2 Related Works 12
• 3.2.1 Unsupervised Person Re-Identification 12
• 3.2.2 Metric Learning with Proxies 14
• 3.2.3 Contrastive Learning and Hard Negative Mining 14
• 3.2.4 Attention-Based Person Re-ID 15
• 3.3 Methodology 16
• 3.3.1 Preliminary 17
• 3.3.2 Hardest Camera Proxy Mining 18
• 3.3.3 False Negative Proxies Mining 21
• 3.3.4 Pyramid Kernel Global Context block 23
• 3.3.5 Optimization 25
• 3.4 Experiments 26
• 3.4.1 Experimental Settings 26
• 3.4.2 Ablation Study 27
• 3.4.3 Parameter Analysis 34
• 3.4.4 Comparison to State-of-The-Art Methods 37
• 4 Auxiliary Attribute Clustering 40
• 4.1 Introduction 40
• 4.2 Related Work 43
• 4.2.1 Unsupervised Person Re-Identification 43
• 4.2.2 Pedestrian Attribute Recognition 44
• 4.3 Methodology 44
• 4.3.1 Problem Formulation and Attribute Information 44
• 4.3.2 Attribute Generation and Label Refinement 46
• 4.3.3 Feasiblity Analysis 47
• 4.3.4 Auxiliary Attribute Clustering Module 50
• 4.4 Experiments 51
• 4.4.1 Dataset and Evaluation Metrics 51
• 4.4.2 Implementation Details 51
• 4.4.3 Comparison to State-of-The-Art Methods 52
• 4.4.4 Ablation Studies 53
• 4.4.5 Discussions and Parameter Analysis 54
• Analysis of the different versions of attribute label 54
• Analysis of the balanced factor α 55
• More Visualization 56
• 4.5 Conclusion 56
• 5 Text-Guided Visual Representation Learning 57
• 5.1 Introduction 57
• 5.2 Related Work 60
• 5.2.1 Person Re-identification 60
• 5.2.2 Vision-Language Model and Prompt Engineering 61
• 5.2.3 Cross-Modal Alignment and Fusion 61
• 5.3 Proposed Network 62
• 5.3.1 Preliminaries 62
• 5.3.2 Cross-Modal Fusion Module 64
• 3.3.3 Optimization Functions 67
• 5.3.4 Overall Network 67
• Training Stage 68
• Inference Stage 68
• 5.4 Implementation configuration 69
• 5.4.1 Datasets and Evaluation Protocols 69
• 5.4.2 Implementation Details 69
• 5.5 Experimental Results 70
• 5.5.1 Comparison to State-of-The-Art Methods 70
• 5.5.2 Ablation Study 71
• 5.5.3 Experiments of hyperparameter 75
• Parameter Analysis on balancing weight of λ 75
• Parameter Analysis on temperature coefficient τ 76
• Parameter Analysis on updating momentum µ 77
• 5.5.4 Qualitative results and discussions 77
• 6 Conclusion 79
• 6.1 Future Work 80
• A Publications 81
• A.1 Journal 81
• A.2 Conference 82
• Bibliography 86