Person Re-Identification Based On Person Appearance Enhancement And Spatial-Temporal Representatives Selection

Person re-identification(Re-ID)is a hot research topic in the field of pattern recognition.The task of person Re-ID is to judge whether the person in the surveillance video scene are the same in a non-overlapping multi-camera network.Therefore,person Re-ID has important application value in the management of the security and transportation department.Of course,it also has broad application prospects in the commercial field.However,person Re-ID is a challenging problem because of the the complexity of application scene,such as background occlusion,illumination in different period,variation in angles and the appearance of person and so on.In recent years,the study of person Re-ID can been divided into two main areas:(1)Person Re-ID based on feature description,the purpose is to design characteristic descriptors which are robust to variation of person appearance and illumination and so on.(2)Person Re-ID based on metric learning,the task is to design a more discriminative distance metric function,which makes the similarity between the same person far greater than the similarity between different person.In view of the above-mentioned challenging problems,this paper has carried out related research on single-shot and multi-shot person Re-ID,as follows:(1)In order to suppress the effects caused by complex background and occlusion,this paper propose a person Re-ID method based on person appearance enhancement model,which use the optimized manifold ranking algorithm to design robust weighted visual saliency features.Specifically,according to the composition relationship,the background query information is initialized to obtain the background enhancement value and the foreground query information,to calculate the foreground enhancement value.In order to optimize feature enhancement,the foreground and background enhancement values are combined to produce the final person appearance enhancement effect.Finally,the appearance enhancement effect and local maximum occurance descriptor are combined to construct a more robust appearance enhanced feature representation.And the model can effectively enhance the person body area and effectively suppress the influence of the background area in the image.(2)For the multi-shot person Re-ID,how to represent the video sequence is an important issue.We observed that there are many similar video frames in the video sequence of each person,which indicate that there are a lot of information redundancy in the research process.Based on the observation,this paper proposed a spatial-temporal representatives selection model to select a representative subset instead of the entire video sequence.Specifically,the subset selection is solved by using a convex function as a row-sparsity regularized minimization problem.In addition,considering the influence of complex background and occlusion,we use the restart random walk model to design robust weighted block descriptor,and successfully perform the person appearance enhancement operation.Finally,the cross-view discriminant analysis is used to reduce the cross-view gap between different cameras.