Action Recognition Of Human Skeleton Based On Spatio-temporal Graph Convolutional Neural Network

Video action recognition has been a highly active research area of computer vision in recent years.Video skeleton data can both provide more compact and accurate information in dynamically changing environments or backgrounds and reduce storage costs and computational costs,so researchers are trying to use skeleton models for action recognition.The introduction of Graph Convolutional Neural Network(GCN)is an effective attempt to achieve the skeleton action recognition,which models the skeleton data as a spatiotemporal graph model with joint points as vertices and bones as edges.However,the complexity and diversity of human motion in real scenes.Traditional GCN can no longer meet the needs of accurately analyzing human action patterns.In addition,the skeleton data has certain limitations,such as noise and occlusion problems.These series of problems make human skeleton action recognition technology face huge challenges.Therefore,in view of some limitations in the traditional methods of skeleton action recognition based on GCN,this paper conducts corresponding innovative research,and its main work is as follows.(1)Action recognition method based on Gaussian noise disturbance.Traditional GCNbased skeleton action recognition usually uses an artificial predefined adjacency matrix to represent the connection relationship between adjacent joint points.This simple predefined structure ignores some important connections existing between non-adjacent joint points and does not integrate well with a self-learning model like neural networks.To address this limitation,perturbation mechanism based on Gaussian noise is proposed to optimize the spatial structure of skeleton data and effectively mining the information between nonadjacent joint points in the spatial structure of the human skeleton.The experimental results on two datasets,NTU RGB+D and Kinetics-Skeleton,show that the method improves the performance of action recognition while saving computational effort.(2)Action recognition method based on two-stream data selection module.The traditional GCN-based skeleton action recognition method ignores the rich spatio-temporal information in the skeleton data,and there are some limitations in the skeleton data itself.Therefore,data selection module is designed,which starts from the three dimensions of the skeleton data,and weights the importance of channels in different dimensions for the characteristics of different skeleton samples,so that the network can effectively focus on more representative features.The inter-frame difference of the skeleton data is also added to the network as a motion stream to supplement the information in the temporal dimension,and then the two-stream data selection module is used to simultaneously mine spatio-temporal information and perform human skeleton action recognition.Extensive experiments on two datasets show that the method can further improve the performance of skeleton action recognition with recognition accuracies as high as 96.14% and 61.0%,reaching the advanced level in the field.