Deep learning models for person re-identification and multi-person tracking

This thesis presents a systematic study of Person Re-Identification (ReID), a core computer vision task enabling the association and re-identification of individuals across different times and camera views. ReID enables applications from multi-camera tracking and video surveillance to non-invasive behavioral analysis, playing a key role in how automated systems balance technical innovation with privacy. The research is structured around four pillars: (RQ1) identification of challenges and trends in supervised and unsupervised ReID; (RQ2) exploration of embedding methods to extract identity-specific features while minimizing transient attributes like pose and background; (RQ3) evaluation of robustness to domain shifts and strategies for domain adaptation; and (RQ4) assessment of integration into multi-person tracking systems. The thesis combines theoretical analysis, empirical evaluation, and proof-of-concept implementations. Experiments with Denoising Diffusion Models demonstrate disentangling stable identity features from transient cues and potential for generative data augmentation. Domain adaptation studies show that carefully selected combinations of heterogeneous datasets improve cross-domain performance, highlighting the importance of data-centric adaptation strategies. Video-based ReID investigations show that Transformer architectures achieve state-of-the-art results by leveraging temporal consistency across frames. An ablation study on ReID integration in multi-person tracking reveals robustness depends on the deployment scenario: high benchmark performance does not ensure reliable real-world tracking. Overall, excelling on benchmarks alone is insufficient. The thesis highlights the importance of integrating discriminative feature extraction, domain adaptation, and system-level design into a coherent framework. Future directions include unsupervised learning, generative augmentation for domain shifts, multimodal pre-training with temporal modeling, and continual adaptation. Ethical considerations are addressed, emphasizing the responsible deployment of ReID technologies in practice.