| Person re-identification is a method to match the target pedestrian in multi camera surveillance image or video from non-overlapping perspective,and it is one of the most active research fields in computer vision.Due to the complexity of the actual monitoring scene,the photographed pedestrian image will appear some problems,such as the change of view angle,the interference of background objects and the uneven space,which brings a great challenge to the research of person re-identification.In order to solve these problems mentioned above,this paper uses a mature single human pose estimation method,Parsing Induced Learner,to estimate the pose keypoints,and obtains the coordinate information of 16 joints of the pedestrian image.According to these joints,the left and right boundaries of the pedestrian image are demarcated,and the pedestrian images of three body areas,namely the upper body,the middle torso and the lower body,are obtained by using six joints which have strong robustness to the camera angle of view.Taking the adjusted Res Net-50 as the backbone network,this paper constructs a simple yet effective Multi-branch Body Region Alignment Network,which consists of two modules: body region extraction and feature learning.The body region extraction module uses the pose keypoints estimated by the pose estimation method,and designs an interception method to obtain three different body region images.In the feature learning module,four global or local branch networks share the base layer,and simultaneously learn the depth features of global and three local body regions.We weighted and fused them to get the final feature expression.In this paper,a large number of comparative experiments and parameter tests are carried out on two large-scale standard data sets Market-1501 and Duke MTMC-Re ID,and the optimal feature fusion strategy is selected to improve the overall performance of pedestrian recognition system.Compared with the same method of pedestrian recognition based on parts in recent years,the proposed method has a great improvement in the evaluation indexes Rank-1 and m AP,which proves the effectiveness and the advanced nature of the method. |
