Research On Person Re-identification In Non-overlapping Surveillance Scenarios

Person re-identification(re-ID)across non-overlapping camera views in surveillance systems indicates that given one image or video of a person in a surveillance network,it could match the corresponding person in the gallery from another camera view.This technique could improve the forensic efficiency.Meanwhile,it could lay a foundation for subsequent advanced monitoring tasks such as suspicious behavior detection,cross-camera people tracking,adverse event prediction and so on.However,person re-identification across non-overlapping camera views in surveillance systems becomes a challenging problem,since the person possesses both rigid and non-rigid properties.Besides,the appearance of the person is easily influenced by illumination,pose,imaging condition,viewpoint changes as well as occlusion.Therefore,how to design an effective and robust algorithm for re-identification attracts the attention of researchers.In this paper,the technique of person re-identification across non-overlapping camera views in surveillance systems is investigated and analyzed.The main achievements contain three aspects as follows:(1)Aimed at inaccurate feature representations caused by indistinct appearance differences in person re-identification,a matrix metric learning for person re-identification based on bidirectional reference set algorithm is put forward.Firstly,reciprocal-neighbor reference sets are obtained by the reciprocal-neighbor scheme.Subsequently,to ensure the robust of references,jointly considering reciprocal-neighbor reference sets from different camera views to generate bidirectional reference sets.Hard samples are mined for feature descriptions via bidirectional reference sets.Finally,these feature descriptors are utilized for matrix metric learning for higher re-identification rate.(2)In order to alleviate highly unstable data distribution due to the data insufficiency,a regularized hull distance learning method with more stringent constraints are proposed.This method firstly represents video features of people as points in the regularized affine space.And then,it expands these feature points to a regularized hull.A regularized nearest point pairs are generated by every two hulls for inter-class and intra-class distance computations.Ultimately,the revised distance model are utilized to top-push distance learning for better person re-identification.(3)For the purpose of excluding many mismatches which are introduced due to purely dependent on visual content relations between query and gallery videos,a Dynamic Hybrid Graph Matching(DHGM)method is designed for unsupervised video-based person re-identification.This method proposes a novel dynamic hybrid model for graph matching which jointly considers content and context information.The model firstly adopts Mahalanobis metric to acquire content dissimilarities.According to the ranking list obtained by content dissimilarities,the bidirectional neighbors are acquired.Subsequently,an expansion operation is conducted for generation of the final bidirectional neighbor sets.And then,the Jaccard metric is used to calculate context dissimilarities.Lastly,content and context dissimilarities are aggregated for more accurate dissimilarity description.