Person Re-identification Based On Two-Stream Network

With the increasing attention to social public security and the development of video collection technology and large-scale data storage technology,the massive video resources generated by a large number of surveillance cameras can be used to analyze historical data.For example,the search and location of criminal suspects,dense crowd monitoring and so on,favorable guarantee the city's public security.In this context,Person re-identification(Person Re-ID)has become a research hot-spot with great value and challenge in the field of computer vision.The person re-identification task aims to give a probe image and quickly find the pedestrian images belonging to the same identity as the pedestrian images to be searched from a large number of gallery images.These images are usually discrete frames taken by different cameras.With the development of deep learning in recent years,person re-identification has made a significant breakthrough,and a variety of methods have been derived.In this paper,we mainly propose two improved person re-identification methods based on two-stream network,and the main research work are as follows:(1)Partial person re-identification method based on two-stream network and sparse reconstruction is proposed.Due to the influence of occlusion,it is impossible to detect the whole body features of pedestrians,the traditional person re-identification method cannot accurately deal with this problem.Therefore,this paper proposes to combine the features of different modes with the sparse reconstruction to match the query and gallery image,so as to realize the partial person re-identification.In this paper,bilinear pooling is used to fuse the features of two different modes of appearance network and pose network to obtain better performance.For matching query and gallery image,the sparse reconstruction process realizes the local pedestrian to recognize the input image of arbitrary size.(2)Person re-identification method based on two-stream network and attention mechanism is proposed.In this paper,the attention mechanism is added on the basis of the two-stream network,and the attention mechanism is combined with the pose through the two-stream network,so as to solve the defect that the attention mechanism will ignore some human body edge parts,so as to improve the performance of the network.At the same time,the middle-level feature is integrated into the high-level feature in the attention branch to improve the expression ability of attention mechanism.This paper conducted experiments on the existing standard partial-body pedestrian datasets Partial-REID,Partial-iLIDS and whole-body pedestrian datasets Market1501,CUHK03,DukeMTMC-reID.The results show that compared with the existing methods,the methods in this paper can get more accurate results.