HDA-Net: Hierarchical Dynamic-Aware Attention with Multi-Head Filtering for Occlusion-Aware Person Re-Identification

Pedestrian recognition in complex scenes faces several challenges, especially in occluded environments, where features often lack discriminability and suffer from computational redundancy. To overcome the difficulties associated with identifying pedestrians in environments with obstructions and intricate visual conditions, this paper proposes a hybrid attention framework consisting of two core modules. The first module, CLS-based Hierarchical Interaction Attention (HIA), enhances feature fusion across shallow, middle, and deep layers of the Transformer encoder by adjusting the attention mechanism, achieving a dynamic balance between local and global features while improving computational efficiency. The second module, Token Filtering based on Statistical Features (TFSF), extracts multi-dimensional statistical features (e.g., mean, standard deviation) and applies a dynamic threshold strategy to filter attention weights. This suppresses background and occlusion interference, focuses on key pedestrian features, and is applied throughout the network to purify feature representation, reduce computation, and improve model robustness and interpretability in complex environments. The experimental findings confirm the proposed approach significantly improves identification precision under occlusion while maintaining high computational efficiency.