Research On Sparse Representation Based Person Re-Identification

The goal of person re-identification(ReID)is to automatically re-identify the same person in another disjointed camera view,which can be widely used for intelligent video surveillance and target tracking.ReID is a challenging task due to the person images often suffer from complex variations in illumination,human pose and occlusion.With the improvements in computer storage capacity,the multi-shot person re-identification has drawn widely attention in recent years.The performance of ReID can be significantly improved by efficiently mining the complementary information between images in the person image sequence with same identity.As the sparse representation is a natural method which can utilize the target pedestrians to reconstruct query pedestrians,it is not only robust to occlusion and image corruption,also can effectively utilize the intrinsic correlation between data,therefore,this paper researches multi-shot person re-identification based on sparse representation.The major research contents of this paper are summarized as follows:(1)Due to the sparse representation model constrained by the l0 regularizer is a non-convex optimization problem,the person re-identification methods based on sparse representation generally exploit the convex l1 regularizer to approaching the l0 regularizer.Although the above problem can be relaxed into a convex optimization problem by using l1 regularizer,the equivalent substitution between them needs to meet the Restricted Isometry Property(RIP).Therefore,this paper exploits a l1/2 regularizer to approaching the l0 regularizer,which has a more relaxed RIP than the convex l1 regularizer.The identical person image sequence in the gallery is regarded as a group,the intra-group structure is constrained by l2 regularizer,the inter-group structure is constrained by l1/2 regularizer,then a person re-identification method based on group sparse representation with mixed l2/l1/2-norm is proposed.In order to further enhance the discriminability of the model,the human body structure constraint is introduced.By divided each person image into several neighboring regions,the adaptive group sparse representation model is constructed for each region and all sparse representation models are incorporated to give the re-identification result,finally.(2)As the person images in the same sequence have similarities and complementary information,compared with single-shot based ReID methods,the multi-shot ReID is more robust to object occlusion and cluttered background.However,the conventional multi-shot ReID methods based on sparse representation only consider the complementary information between the images in the same person image sequence in the gallery and ignore the prior information contained in the same person image sequence in the probe,i.e.usually encode each probe person image individually.Therefore,we propose a joint group sparse representation(JGSR)method to simultaneously encode a sequence of probe person images with the same identity.Additionally,the JGSR method can effectively exploit the prior knowledge contained in each of the person image sequence in the gallery and probe sets simultaneously.Therefore,the proposed method is more robust to human pose variations and object occlusion.(3)Person images in the same sequences are similar in appearance.Information is redundant.To address this problem,in this paper,k-means algorithm is employed to generate the clusters of each person sequence.After that,the mean of the image features in each cluster is used as the representative.Additionally,a Kernel Local Fisher Discrimination Analysis(KLFDA)method is introduced to mitigate the cross-view gaps on person re-identification.Finally,we incorporate sparse representation method with k-means clustering and KLFDA metric learning method into a unified framework,resulting in superior person re-identification performance.Extensive experiments on three publicly available iLIDS-VID,PRID 2011,and SAIVT-SoftBio multishot benchmark databases were conducted and demonstrated the superior performance of the proposed methods in comparison with current methods.