Research On Cross-Domain Person Re-Identification Algorithm Based On Graph Convolution Association Learning

In recent years,with the rapid development of deep learning technology and the urgent need to process massive monitoring data,person re-identification has achieved excellent performance in single domain,while also making some good progress in cross domain.However,due to factors such as person posture changes,camera angle of view and resolution differences,occlusion,and lighting,especially in cross-domain scenes,there are also serious background differences that make person re-identification still face many difficulties.This thesis will focus on cross-domain issues,focusing on the core issue of how to reduce domain differences in cross-domain scenarios.This thesis will systematically and comprehensively analyze and study some existing methods from three aspects: style transfer,discriminant feature extraction,and unsupervised domain adaptation,and fully learn the intra and inter class correlations and inter-domain correlations of target sample features using graph convolution related techniques,and committed to training cross-domain models with strong robustness.This thesis adopts a gradually in-depth research approach,from domain generalization to domain adaptation,which is divided into two stages: source domain pre-training and target domain optimization.The main innovation points are:(1)By improving the introduced attention module and building a person background conversion unit based on Cycle GAN,the network’s attention ability to pedestrian foreground and background is enhanced,and a constraint loss is designed to control the network’s generation of high-quality intermediate samples to enhance correct learning of the network,achieving the ideal effect of background conversion.(2)By integrating the introduced local feature comparison module and graph convolution module,a graph convolution feature fusion unit is designed.Res Net-50 is used as the backbone network to extract global,local relationship,posture relationship,and topological features of persons,resulting in a complete representation of persons with rich granularity and relevance information,effectively improving the robustness and generalization ability of the model.(3)By integrating the introduced dual graph convolution and feature adaptation module,a graph convolution domain adaptation unit is constructed.On the one hand,dual graph convolution operations are performed on the fused features after clustering,enabling the model to effectively propagate identity information within and between classes.On the other hand,inter-domain distribution alignment loss is constructed for the clustered posture features,fully exploring domain invariance,in order to cyclically update the pseudo labels and the entire network.This greatly improves the domain adaptation ability and recognition accuracy of the model.A large number of ablation comparative experiments on Market-1501 and DuckMTMC-reID datasets have verified the effectiveness of the methods in this thesis.The final complete algorithm in this thesis can achieve better recognition performance in some cross-domain scenarios,and has certain engineering application value.