A dual-branch pedestrian re-identification method CPHMNet based on multi-dimensional feature fusion and integrated pose estimation

To address the interferences in pedestrian feature extraction caused by pose variations and occlusions, which adversely affect recognition performance in pedestrian re-identification tasks, this paper proposes a pedestrian re-identification method, CPHMNet, based on multi-dimensional feature fusion and integrated pose estimation. Firstly, a novel MDA module is developed to enhance the capability of feature discrimination for similar pedestrians. The model employs a dual-branch network architecture, where the first branch processes through the pose estimation branch to capture body posture and local features; The second branch introduces the CBAM attention mechanism to boost the model's proficiency in extracting information. The two branches are weighted and fused by the designed multi-dimensional feature fusion network (MDFF), which considerably optimizes the model's performance by capturing global information with different dimensions and enhancing features. Finally, a joint loss function is introduced to impose constraints on the model. Substantial experiments on the Market1501, DukeMTMC-reID, and MSMT17 datasets demonstrate that the mAP values reached 89.9%, 81.1%, and 61.6%, respectively, while the Rank-1 accuracies achieved 95.9%, 90.3%, and 82.5%. This method displays superiority in satisfying the practical demands of pedestrian re-identification in real-world scenarios.