Person Re-identification Methods Based On Semantic Correspondence And Muti-view Matching

The task of Person Re-identification(Person Re-ID)is the key to the modern city-video surveillance systems.The person tracking technology and person retrieval technology have the wide application scenarios and the extensive and profound significance for the society,especially,for the maintenance of public order and the improvement of the efficiency of criminal inves-tigation.Person Re-ID is essentially the problem of person images matching,which face to the challenges from the variations of the vision appearance caused by viewpoint changes from such as pose,luminance,background and so on.To deal with these challenges above,this paper presents our works from two aspects:robust image representation learning and effective matching strate-gies learning.On one hand,the learned image representation is able to represent the intrinsic information of images.On the other hand,the learned image representation is the comprehensive representation for the images.The learned matching strategies are to maximum the gap between the positive samples and negative samples.Based on the analysis above and deep learning,three algorithm frameworks are proposed,which include the Pyramid Person Matching Network based on the pyramid matching module,the Pyramid Person Matching Network based on the color-texture distribution and the Pyramid Person Matching Network based on multi-channel features,which is the improved methods of the former two methods.The extensive experiments on the pop-ular datasets demonstrate the effectiveness of the proposed methods and also show the advance performance,compared to the state-of-art methods.