Person re-identification based on random occlusion and multi-granularity feature fusion

Aiming at the problems of occlusion and monotony of pedestrian discriminative feature hierarchy in person re-identification, this paper proposes a method combining random occlusion and multi-granularity feature fusion based on the IBN-Net50-a network. First, in order to enhance the robustness against occlusion, random occlusion processing is performed on the input images to simulate the real scene of pedestrians being occluded. Secondly, the network includes a global branch, a local coarse-grained fusion branch and a local fine-grained fusion branch, which can extract global salient features while supplementing local multi-grained deep features, enriching the hierarchy of pedestrian discrimination features. Furthermore, further mining the correlation between local multi-granularity features for deeper fusion. Finally, the label smoothing loss and triplet loss jointly train the network. Comparing the proposed method with current state-of-the-art person re-identification algorithms on three standard public datasets and one occlusion dataset. The experimental results show that the Rank-1 of the proposed algorithm on Market1501, DukeMTMC-reID and CUHK03 is 95.2%, 89.2% and 80.1%, respectively. In Occluded-Duke dataset, Rank-1 and mAP achieved 60.6% and 51.6%. The experimental results are better than those of the compared methods, which fully confirm the effectiveness of the proposed method. © 2023 Beijing University of Aeronautics and Astronautics (BUAA). All rights reserved.