Pedestrian re-recognition method based on dual attention and GRU

Abstract

In view of the existing problems of pedestrian re-recognition in real scenes, such as camera Angle change, pedestrian posture change and object occlusion, in order to improve the accuracy of pedestrian re-recognition and enhance the ability of pedestrian feature representation, a pedestrian re-recognition method combining double attention and gated cycle unit (GRU) was established based on ResNet-50. The global structure information of pedestrian images can be better captured by channel attention and spatial attention, and the temporal information of pedestrian images can be obtained by GRU. Combined with the triad loss training network of difficult sample sampling and the cross-entropy loss training network based on label smoothing, the experiments were carried out on dataset Market1501 and CUHK03, and the accuracy of rank-1 reached 96.2% and 76.7%, respectively, which increased by 2% and 2.9% compared with the original ResNet-50 backbone network. Experimental results show that the proposed model can effectively improve the accuracy of pedestrian re-recognition.