Cross-Modality Person Re-Identification: An Attention-Enhanced Framework for Deep Fusion of Visible and Infrared Features

The technology of cross-modality person re-identification using visible light and infrared has gained widespread attention due to its ability to enable all-weather intelligent monitoring, unaffected by nighttime and low-light conditions. A major direction in current research on cross-modality person re-identification is to reduce the visual feature discrepancies between the two modalities, including variations in illumination, pedestrian disguise, and shadow effects. To address this issue, we propose a network structure based on dual-modal feature fusion, focusing on automatically discovering and enhancing the shared features between the two modalities through an attention mechanism while suppressing the differing features. Additionally, a dual-branch network structure is employed to separately extract features from visible light and infrared images, with a cross-attention module facilitating feature fusion, thereby yielding more robust pedestrian representations. The proposed method demonstrates ideal performance on mainstream datasets SYSU-MM01 and RegDB, showcasing its strong adaptability to various imaging conditions and providing effective references for future multi-modal learning research.