Research on Re-recognition Method of Multi-branch Fusion Attention Mechanism for Occluded Pedestrian

Due to the problem of feature misalignment between pedestrian images, illumination, posture, occlusion and other factors have a great impact on re-recognition, an attention-based multi-branch fusion pedestrian re-recognition framework is proposed. Firstly, Resnet50 is used as the backbone network to extract the initial features, and then the extracted initial features are entered into the network in parallel, and the final features are obtained through feature fusion. The network was trained by cross entropy loss, and experiments were performed on Occluded-ReID data sets, P-ETHZ data sets and Partial-ReID data sets with serious occlusion. The accuracy was significantly improved, which proved that this method could improve the recognition rate in practical application scenarios.