Pedestrian Re-identification Method Based On Block Occlusion And Space Alignment

Person re-identification is an important research direction in the field of computer vision and belongs to a branch of image retrieval.The main task is to use visual algorithm(deep learning method)to pair the pedestrian images or videos extracted under several non-overlapping and non-crossing devices,that is,when given a pedestrian image,to retrieve the pedestrian images in the cross-region video surveillance system.This direction can be widely used in video surveillance,criminal investigation,intelligent security and other social fields.According to the different types of methods,pedestrian re-identification methods can be divided into two categories: pedestrian re-identification based on non-deep learning and pedestrian re-identification based on deep learning.With the rapid development of deep convolutional neural networks,research on pedestrian re-identification methods based on cross-modal,attention mechanism,attitude estimation,super resolution and unsupervised has been vigorously developed.However,pedestrian re-identification still faces many challenges,such as effective training of large-scale data,robustness under changes in pedestrian appearance,and application under unrestricted scenarios(occlusion,illumination,etc.).In order to solve the above problems,a Batch Part-mask Network and a Pedestrian Alignment Batch Part-mask Network are proposed in this paper.1)Batch Part-mask Network is composed of global feature branch and feature occlusion branch,with Res Net-50 as the backbone network.The global branch is used to learn and encode the global features,and the batch part-mask branch composed of the first branch and the second branch is a double-branch structure with feature occlusion function to learn and encode the local detail features.Let the feature tensor of a single batch of input images calculated by the backbone network be T.The batch erasure layer in the first branch randomly erases the same region of the tensor,returning all values in the erasure region to 0.The second branch will first divide the input feature map into upper and lower two pieces,and then randomly block a small piece in each piece,that is,all values in the blocked area are set to 0.The feature tensors obtained by the erasure of the first and second branches on the feature graph are T? and T? respectively.In general,the erasure area should be large enough to cover the semantic portion of the input feature graph.The goal is to learn multiple feature areas of interest,rather than focusing on the main recognition area.2)Pedestrian alignment network is based on Batch Part-mask Network and two modules are added to global branch,which are mapping estimation network and alignment branch.The mapping estimation branch is used to realize the spatial transformation function.The fourth convolution module of Res Net outputs the feature map with the size of 14*14*1024 as the input,the input mapping estimation network carries out the convolution transformation,and the output vector of 6dimensions is used to generate the image grid,so as to solve the problem of too many backgrounds and partial missing.The Alignment branch structure is similar to the backbone network and contains the Res Net 3 convolution module and Res Net 4 convolution module.For the input image,we obtain pedestrian features from the original image and the aligned image.We obtain pedestrian features from the original image and the aligned image.Better alignment of pedestrians can learn more discriminant features and improve the accuracy of pedestrian matching.The accuracy of Rank-1 on the Duke dataset is improved by 1%-2%,reaching 87.1%.The accuracy of m AP on Market-1501 data set is improved by 2%-3%,reaching 86.4%.The test results on several large-scale datasets show that the proposed Batch Part-mask Network has good robustness and improves the accuracy of pedestrian detection in multiple video surveillance.The Pedestrian Alignment Batch Part-mask Network based on block occlusion can solve the problem of excessive background and partial missing of the input pedestrian image under the condition of strong ability to extract local features,so that the input pedestrian image can achieve semantic alignment and enhance the ability to extract global information.The proposed two methods have achieved good results in pedestrian re-identification.