Research On Double-Branch Person Re-Identification Method Based On Graph Convolutional Networks

Person re-identification aims to accurately identify and match the same pedestrian across different surveillance cameras,and has broad application prospects in fields such as urban security monitoring,intelligent transportation,and crowd counting.However,there are still urgent problems that need to be addressed.Firstly,images captured by surveillance cameras in real scenes often suffer from problems such as lighting changes and occlusion,and existing methods mostly rely on convolutional neural networks,which have two limitations: limited local receptive fields and possible loss of important fine-grained information due to downsampling.Secondly,although visual Transformers overcome the two limitations of convolutional neural networks,their ability to extract local detailed features is relatively weak.To address these issues,the main contributions of this paper are summarized as follows:(1)Aiming at lighting changes,occlusion,and two limitations of convolutional neural networks,this paper proposes a global-local dual-branch person re-identification method based on graph convolutional network and visual Transformer improvement.Firstly,multi-modal data augmentation is used to fit the spatial and color information of visible and grayscale modal images in a shared space,reducing interference from lighting changes.Secondly,the Transformer is used to replace the convolutional neural network,and auxiliary information embedding is introduced to fuse non-visual information,reducing feature bias caused by camera changes.Finally,a global-local dual-branch architecture is designed to simultaneously learn robust global features and discriminative local features.The local feature branch introduces graph convolutional networks,where the GCN and Transformer layers form the GAT module.The GCN explores the relationships between nodes to learn discriminative local features within image slices,while the Transformer learns robust global information between image slices.The combination of the two can effectively learn discriminative pedestrian features.Experimental results on three mainstream datasets show that the proposed method can effectively improve the performance of person re-identification.(2)Aiming at the weak extraction ability of the local detail features of the visual Transformer,this paper proposes a dual-branch person re-identification method based on graph convolutional network and visual transformer improved by convolutional attention.This method mainly embeds the channel attention and spatial attention in the convolutional layer into the model architecture to strengthen the focus on key areas and important channel features in the image,thereby further improving Transformer’s ability to extract local detail features.Experimental results on three mainstream datasets show that this method can effectively improve the model’s ability to extract local detail features under occlusion.Finally,using the Pycharm integrated development platform,the Python programming language,and the Pytorch framework,and with the help of the cross-platform toolkit PyQt5 for graphical interface design,a simple pedestrian re-identification software was developed.