Research On Person Re-Identification Method Based On Attention Mechanism

The goal of the Person re-identification problem is to find a specific pedestrian under the condition of cross-camera,which is widely used in intelligent security,smart city and other major national needs.As a kind of image recognition tasks,it uses the pedestrian images collected from different camera as the training set.With the influence of pedestrian properties,environment,posture,shade and camera resolution,including the difficulty of make training set,person re-identification method still has great space for development.Attentional mechanism method can make the model focus on the key part of the image and suppress the interference of the irrelevant part,which is suitable for the person re-identification problems.Some previous person re-identification methods based on attention mechanism often use global pooling or convolution with finite receptive field to learn,which is difficult to mine the correlation between pedestrian feature points in fine granularity.In view of the above problems,this paper mainly carries out the following work:(1)Based on the previous related work of attention mechanism and feature alignment,this paper proposes a feature segmentation module based on the selfattention mechanism,it can make the model mining the correlation between image feature points in each pedestrian part in fine granularity,then adaptively learn the attention map according to the correlation,which makes the model focuses on learning key features while suppressing irrelevant features.(2)This paper explores the effectiveness of using this method separately and jointly in spatial dimension and channel dimension as well as the effectiveness of feature segmentation operation.Experiments on market-1501 and Duke-MTMC,two important person re-identification datasets in the academic world,show that:Compared with the original ResNet-50 network(baseline),the proposed method improves the rank-1 and mAP indexes by 2.1%and 4.8%,respectively.In addition,compared with the previous methods based on feature alignment and attention mechanism,the proposed method has different degrees of improvement in recognition accuracy,and has good robustness against occlusion and image noise.