Gait Recognition Research Based On Multi-Stream Feature Fusion Of Skeleton And Graph Convolution

As the public pays more and more attention to the safety of public areas,gait recognition as a biometric technology has also attracted more and more researchers’ attention.Compared with other biometric technologies,gait recognition can realize long-distance recognition,and has the advantages of being difficult to camouflage,non-contact and strong security,which makes it show important research significance and research prospects in the fields of security and criminal investigation.In view of the inherent robustness to complex scene changes in the dynamic nature of the human skeleton model and the excellent performance and generalization ability of the graph neural network for such non-European structures.In this paper,a gait recognition method is studied based on 2D human skeleton model and spatio-temporal graph neural network.The traditional gait recognition method based on skeleton model usually regards the joint information in each frame as a separate feature vector,ignoring the spatio-temporal connection between joints.This paper proposes a spatio-temporal graph convolution-based gait recognition algorithm Gait St.The features of human skeleton are preprocessed by data augmentation method,and the spatio-temporal topology map of human skeleton is constructed,taking into account the spatio-temporal connection between joints.The motion representation of the joint neighborhood is divided by the spatial configuration partitioning strategy,and the motion of the local part is effectively modeled.High-dimensional features are sequentially extracted from spatial and temporal dimensions through spatio-temporal graph convolution and aggregated to obtain highdimensional representations of skeleton features.In order to improve the robustness of the algorithm to complex environmental changes,this paper proposes an improved gait algorithm Gait Eff based on multi-stream feature fusion and channel attention spatio-temporal graph convolution based on Gait St.By extending the skeleton features to form multi-stream branches and adopting an early fusion structure,the model fuses more effective gait features.The feature input layer is improved by channel attention spatiotemporal graph convolution,so that the model can learn the topological relationship of different channels representing different motion states,breaking the modeling limitation of the original graph convolution.The temporal convolution is improved by separable convolution,which improves the feature extraction capability of mainstream networks and reduces model complexity.This paper uses the CASIA-B data set published by the Institute of Automation,Chinese Academy of Sciences to conduct multi-state experiments,and analyzes the experimental results through the objective evaluation index Rank-1.The improved algorithm proposed in this paper effectively improves the model’s sensitivity to changes in perspective,carried objects and clothing.The robustness verifies that the algorithm proposed in this paper is effective.