Occluded Person Re-identification Based On Region Division And Deduction

Person re-identification,as a sub-problem of image retrieval,can be widely used in intelligent video surveillance,intelligent security and other fields.Due to the coexistence of camera differences,rigidity and flexibility,and the appearance is susceptible to wear,scale,occlusion,attitude,and perspective,person re-identification has become one of the research hot spots in the field of computer vision that is both academic and challenging.previously re-identify the person re-identification occlusion is one of the difficult points in the research of person re-identification.The dissertation mainly studies the occluded person re-identification based on region division and deduction.Firstly,by previously re-identifying the person re-identification data sets,images that are obviously not the same person are eliminated.Secondly,the pose estimation algorithm is used to divide the person sub-regions as a priori,and the occluded regions are deduced.The main work and research results of the paper are as follows:(1)An occluded person re-identification data set is constructed.According to the current situation that the existing occluded data set does not have complete and occluded images,the existing conventional person re-identification data set is used to construct the occluded person re-identification data set.(2)Advance person re-identification is used to eliminate images that are obviously not the same person,improving the overall recognition efficiency.Advance person re-identification was performed using simple person sub-region division and HSV color histogram extraction methods.Various image similarity calculation methods are compared,and the Bhattacharyya distance is finally selected for image similarity calculation.(3)An occluded region deduction criterion and method based on person sub-regions division is proposed.Using the pose estimation algorithm as a priori,the person skeleton and key feature points are obtained.Based on this and thecharacteristics of the occluded person images,the person sub-regions are divided,and the occluded sub-regions are deduced according to the criteria.(4)Design dual-stream network FORN to deal with occluded person re-identification.FORN is divided into two parts,FPN and FSN networks.The FPN network extracts features from person images and each sub-regions,and finally performs tree-like feature fusion through a fully connected layer.The FSN network uses the method of sparse feature reconstruction to obtain the similarity between images,and finally obtains the result of occluded person re-identification through the feature fusion calculation method.Experimental results on public data sets Duke MTMC-re ID and Market-1501 and self-built data sets show that the methods in this paper are better than the existing mainstream person re-identification methods.