Multi-Modality Person Re-identification Based On Deep Learning

With the development of public surveillance,the analysis and application of surveillance data have become an important academic research in the smart city.Given a person image,the goal of person re-identification(person re-id)is to retrieval the same person images across non-overlapping cameras,which has a wide application in public surveillance,e.g.,the person re-tracking,the analysis of person behaviour.However,there is a gap between the person re-id research and the real-world application.On the one hand,most of the existing methods mainly focus on the single-modality person re-id task of day-time visible images,and the cross-modality person re-id task between the day-time visible images and the night-time infrared images is largely ignored.On the other hand,the day-time visible images suffer from the human pose,viewpoints,illuminations and other imaging factors,limiting the performance of single-modality person re-id task.This thesis focuses on the cross-modality person re-id and single-modality person re-id tasks,which mainly includes two aspects:For cross-modality person re-id task,this thesis proposes a cross-modality method based on the Semantic Coupling and the Identity-Consistence Constraint.(1)A bi-directional Semantic Coupling module is proposed to alleviate the modality discrepancy.The Semantic Coupling module couples the semantic information of the visible images and infrared images belonging to the same person,enriching the single-modality images with the semantic information of the other modality images.With this module,the modality discrepancy is alleviated and the modality-special information is reserved.(2)The Identity-Consistence Constraint is proposed to alleviate the intra-class variations.Specifically,the cross-modality triplet loss and identity loss enforce the constraint in feature space and distance space,respectively.Extensive experiments demonstrated that the method effectively alleviates the modality discrepancy and the intra-class variations,greatly improving the accuracies of cross-modality person re-id task.For single-modality person re-id task,this thesis proposes a person re-id method based on human correlation graph structure.(1)A fine-grained feature-extracting module is designed: The human feature maps are divided into several body blocks horizontally to extract the fine-grained features.(2)A human correlation graph is built up: The body region blocks are regarded as the graph nodes and the similarities between body region blocks are reviewed as the graph edges,building up the human correlation graph.Finally,the global human structure information is mined from the human correlation graph,which greatly alleviates the inference of human pose,viewpoints,illuminations and other imaging factors.Extensive experiments demonstrated the effectiveness of the method.