Pedestrian Re-identification By Graph Clustering

It is a challenging task to implement pedestrian re-identification in multi-video surveillance networks.Some of the current research results have achieved retrieval of specific pedestrian targets,but the sorted resultreturned by similaritycan't indicate which retrieval results are from the same pedestrian with selected pedestrian target automatically.In this paper,we propose a framework to realize multi-video pedestrian re-identification by graph clustering,We cluster the observed images of all pedestrian objects in themulti-camera surveillance scene,and get the activity trajectory of each pedestrian object in the whole scene.The multi-video pedestrian re-identification framework proposed in this paper includes: feature representation of pedestrian images;transformation of similarity graph of pedestrian images,partitioning the similarity matrix by graph clustering,realizes the clustering of pedestrian objects.Firstly,This paper study the transformation of the feature representation of images based on dColorSIFT feature.Because of the deficiency of the traditional metric method in similarity measure,we introduces the partitioning score(score_ocsvm)technique.At the same time,in order to improve the information quantity of feature representation of images,we propose a method that remove the image background area and integrate image with the sequence image of a period of time,and the ability of image similarity measurement is improved effectively.Then,we study the construction of similarity graph of pedestrian images.This paper achieve the method based on neighbors and points out the boundedness of other methods.We also study the problem of number estimation in graph clustering,Graphs clustering is used to replace the graph partitioning in GN algorithm,and the modularity is used to measure the results.The experiments verify its' feasibility.Finally,pedestrian re-identification experiments are carried out on standard dataset and self-construct dataset,and the multi-video pedestrian re-identification framework proposed in this paper is verified.Especially for the self-construct dataset which aim at practical application,we analysis the actual difficulties and put forward relevant solutions to improve the pedestrian re-identification effect.