POSE-GUIDED MULTI-PART ATTENTION FRAMEWORK FOR VIDEO-BASED PERSON REIDENTIFICATION

Person Re-Identification (ReID) is a fundamental task in intelligent video surveillance, aiming to recognize individuals across non-overlapping camera views. Despite significant progress, video-based ReID remains challenging due to variations in illumination, occlusion, and dynamic background clutter. To address these challenges, we present a novel ReID framework that jointly exploits spatial and temporal cues for robust person retrieval. The proposed approach integrates multi-level spatial representations by combining global, intermediate, and fine-grained features. Global and intermediate information is captured through self-attention mechanism, while fine-grained details are derived from pose-based attributes to enhance discriminative power. These spatial representations are then fused and processed by a temporal self-attention module, which models inter-frame dependencies to effectively capture motion dynamics. A retrieval stage subsequently leverages the joint spatial-temporal features to achieve accurate identification. Comprehensive experiments on the PRID2011 and iLIDS-VID benchmark datasets demonstrate the effectiveness of the proposed method. The framework achieves consistent improvements over baseline approaches, with gains of 2.3% and 3.4% in mean Average Precision (mAP) on PRID2011 and iLIDS-VID, respectively. These results highlight the robustness and discriminative capability of the proposed architecture for real-world video-based ReID applications.