Pedestrian re-identification algorithm based on global attention and region ranking

Feng-Ping An, Jianrong Wang, Ruijun Liu

Abstract: —In practical applications, significant variations in camera parameters and the complexity of shooting environments present substantial challenges for pedestrian re-identification. These challenges include inconsistencies in pedestrian image features across diverse scenes and a limited availability of difficult instances, such as occluded or low-resolution images. To address these issues, we propose a global attention mechanism that fully considers the temporal and spatial consistency of pedestrian information. This mechanism effectively captures regional features that best characterize pedestrian attributes while ensuring that these regions remain temporally and spatially continuous. As a result, robust feature representation is maintained across different scenes, even under occlusion. Furthermore, to overcome difficulties in sample annotation and the ambiguity of boundaries within deep learning-based pedestrian re-identification algorithms, we propose a pedestrian recognition algorithm based on salient region ranking. This algorithm reformulates the identification of salient regions within pedestrian images as a multi-label ranking task, employing ordinal regression to uncover correlations among different category labels, thereby enhancing the robustness of the region saliency classifier. Ablation and comparative experiments conducted on the Market1501, DukeMTMC-reID, and Occluded-Duke datasets demonstrate that our method not only achieves superior recognition performance but also validates its high generalization capability and robustness. This research offers significant theoretical and practical contributions to the further development of deep learning techniques in pedestrian re-identification.