Research On Gait Recognition Combined With Human Pose And Graph Convolution

Gait is an important biometric feature used to identify people,attracting more and more attention in the field of video surveillance because of its advantage of being able to capture features at a distance and without cooperation.Compared with the silhouette-based method,the application of human pose can effectively alleviate the influence of covariate changes on gait recognition performance.However,existing pose-based approach tend to model skeleton data as a vector sequence or pseudo image,and then send it to CNN or RNN for processing.due to the lack of effective use of information on joint points,the recognition effect is not good enough.Considering that the human skeleton topology graph has good stability and invariance,we propose to introduce the concept of graph into the gait recognition task.The main points of innovation and contributions are as follows:(1)A partitioned spatial temporal graph convolution network for gait recognition is proposed,and a spatiotemporal graph is constructed from joint points sequences to dynamically model skeleton sequences.The model can automatically learn the spatial features and temporal information of human skeleton during walking,and the graph is divided into blocks to learn the high-level attributes between different parts and the relationships between them to achieve fusion of local and global information.(2)In order to extract more distinguishing features,the use of attention mechanisms is proposed to enhance the significance of the extracted spatiotemporal features.What is more,a multi-loss strategy is proposed to optimize the network,reduce the inter-gait distance of the eigenvalues and increase the intra-gait distinction.(3)We constructed a human skeleton dataset for gait recognition based on commonly used gait datasets,based on which experimental comparisons were performed on the cross-view recognition task and the recognition task without view variation in perspective,and the effectiveness of the model is verified through the ablation experiments.Experiments have shown that this method is not only robust to clothing and carrying states,but also has good generalization ability on cross-view tasks.