Efficient person re-identification via progressive filter pruning and body part-aware feature learning

Person re-identification (Re-ID) is a critical task in modern surveillance, enabling the tracking of individuals across multiple camera views despite variations in appearance, pose, and background. As public safety demands grow, Re-ID systems must operate efficiently in real-time, especially on resource-constrained devices. However, traditional deep learning models are often too computationally intensive for practical deployment. This research presents an optimized Re-ID framework that integrates Progressive Soft Filter Pruning (PSFP) with body part-aware local feature learning. PSFP reduces model complexity while limiting accuracy loss, and local feature learning enhances robustness against occlusions and appearance variations. Extensive evaluations on benchmark datasets demonstrate that the proposed full BPBreID model achieves 92.19% Rank-1 accuracy on Market-1501, while the pruned BPBreID + PSFP model retains 84.61% Rank-1 accuracy with a 37% reduction in FLOPs and 16.7% decrease in memory usage. These results highlight the trade-off between computational efficiency and recognition performance, confirming the feasibility of lightweight, scalable Re-ID solutions for real-world deployment. The complete implementation is available at [{https://github.com/women-ssniffp/Person-ReID-PSFP.git].