Person re-identification (Re-ID) evaluates individual consistency across surveillance cameras and is crucial in intelligent video analysis. This study introduces three innovations: (1) Keypoint-Based Dynamic Region Partitioning (KDRP), ad-dressing limitations in traditional stripe segmentation under varying angles and postures; (2) Adaptive Average Pooling Layer List Network (APLNet), enhancing feature extraction stability and computational efficiency; and (3) Priority Circle Loss (P-Circle Loss), assigning higher weights to key samples to boost model discrimination and accelerate training. Ex-periments on the filtered Market-1501 dataset achieved 85.61% mAP and 96.42% Rank-1 accuracy, showcasing superior adaptability to variations and improved Re-ID accuracy.