Research On Cross-perspective Pedestrian Re-identification Based On Dictionary Learning

Person re-identification,as one of the key tasks of intelligent video surveillance analysis,can automatically match pedestrian images from across camera views.Person re-identification has become a key technology in the field of computer vision and machine learning.It is precisely because of the wide application prospects of pedestrian recognition technology in practice that person re-identification has become a research hotspot in the field of computer vision.Although a lot of research progress has been made,due to the complexity of camera surveillance scene,pedestrian recognition technology is still facing great challenges.In practice,due to the differences and economic factors of different monitoring areas,different cameras are discontinuous,and the monitoring areas do not overlap with each other.Therefore,this will make it difficult for pedestrian recognition tasks.For instance,(1)the same pedestrian image under different camera views has some problems,such as pedestrian posture,background clutter,occlusion and so on,which makes the appearance characteristics of pedestrian image show greater ambiguity.(2)Most existing person re-identification methods pay more attention to the similarity of the same pedestrian in different views,ignoring the influence of similarity components between different pedestrians in different perspectives on the recognition algorithm.However,these components are the main embodiment of pedestrian similarity and the root reason for reducing the recognition rate.In this paper,as to mention above the problem,some research results are as follows:(1)In order to solve the problems in(1),we proposes top distance regularized projection and dictionary learning for person re-identification.By mapping the sparse coding coefficient of the dictionary to a discriminant space,the distance between the same pedestrians under different camera views becomes closer and the discriminant ability of the dictionary is improved at the same time.In addition,all similar distance pairs in training samples are obtained by listwise.Finally,a top distance regularization term is introduced into the framework of joint learning of dictionary and projection matrix,which improves the discriminant ability of dictionary and projection matrix while optimizing the solution space of objective function.Considering the actual situation,the dictionary learning model is extended to multiple perspectives.(2)In order to solve the problem that the similarity of pedestrians under different camera views reduces the recognition performance,a common and specificity dictionary joint learning framework for person re-identification is proposed in this paper.Because each pedestrian is made up of a shared component reflecting its similarity and a unique component reflecting its identity.Therefore,we propose to reduce the ambiguity between pedestrian by eliminating the shared components of features.Therefore,we proposes a joint learning framework of person-shared and personspecific component dictionaries,and introduces the distance and coherence constraints of coding coefficients of the same pedestrian-specific component under the special dictionary,forcing the same pedestrian to have similar coding coefficients,and different pedestrians have weak correlation.In addition,for the shared dictionary and specificity dictionary,low-rank and sparse constraints are introduced respectively to improve dictionary ability and discrimination.