A Study Of Person Re-identification Based On Discriminative Dictionary Learning

In recent years,person re-identification has attracted more and more researchers’ attention because of its extensive application in person search,person tracking and person behavior analysis.Although researchers have made great efforts to enhance the performance of person re-identification,due to the differences between different cameras,and person appearance is vulnerable to clothes,light,and occlusion,pose and perspective make person re-identification technology still a huge challenge.As an important branch of intelligent surveillance system,research on robust person re-identification models and algorithms has high theoretical value.This thesis has carried out a series of studies to improve the matching rate of person re-identification.The main work includes the following two parts:(1)In order to reduce the differences caused by the background,illumination and occlusion between different camera views,a person re-identification algorithm based on the identity consistency and irrelevant constraints is proposed.The algorithm learns dictionaries to represent person image for each view,and different identifiers are used for each view to realize the transform of person visual feature to identity information space.Specifically,we introduce identity consistency and irrelevant constraints to keep the identity information between different persons at a certain distance,and keep the identity information of the same person as close as possible.Finally,similarity measurement scheme of person re-identification is achieved by identity information.Experiments on four datasets,VIPe R,GRID,PRID2011 and CUHK01,show higher recognition accuracy of the person re-identification algorithm after identity information constraint against other methods.(2)Aiming at the problem of domain shift between different camera views,a novel domain-common and domain-invariant dictionary pair learning method is proposed for cross-view person re-identification.Specifically,based on the fact that pedestrians from the same camera view share the same domain,and that the information of domain carried by each pedestrian image in the same view has consistency for short time,we propose to decompose person image from the same view into specific perspective domain information components and domain-invariant pedestrian appearance feature components,and develop a discriminative dictionary learning model to create a domain-common dictionary for characterizing the domain information components and a domain-invariant dictionary for characterizing the pedestrian appearance components.Since images from the same camera view have domain similarity,dictionary used to represent domain information is refined by low-rank regularization.To further improve discrimination ability of learned dictionaries,we propose certain constraints in the algorithm,that is,the coding coefficients of multiple images with the same identity and view to have strong similarity.Furthermore,to overcome the appearance ambiguity,a novel expansion regularization is used to solve the visual appearance ambiguity of similar appearance features of different pedestrians and different appearance features of the same pedestrian.Experiments performed on four challenging datasets demonstrate the effectiveness and the superiority of our approach against some state-of-the-art methods.