STFANet: Learning Spatio-Temporal Feature Aggregation for Person Re-Identification

Identifying individuals across video sequences, known as video-based person re-identification (re-ID), is a crucial task in computer vision with diverse applications such as surveillance, behavior analysis, and smart city management. Unlike methods that rely on single images, video-based re-ID leverages temporal information to improve identification accuracy by capturing motion patterns and spatial-temporal interactions. Despite significant advancements, existing methods often struggle to efficiently integrate motion features while maintaining computational efficiency. This paper presents a novel framework called Spatio-Temporal Feature Aggregation Network (STFANet) that addresses these challenges. It consists of Local Motion Feature Learning (LMFL), which enhances frame-level feature extraction by learning from consecutive frames, and a Global Motion Feature Learning (GMFL) module, which captures comprehensive motion patterns across longer sequences. To assess the effectiveness of the proposed STFANet, experiments are conducted on two well-known datasets: LS-VID and MARS. Experimental results demonstrate that the proposed framework outperforms state-of-the-art models by achieving higher accuracy and improved computational efficiency.