Research On Human Action Recognition Method Based On 3D Skeleton Information Feature

Human action recognition is a complex task.According to different data sources,human action recognition can be divided into human action recognition algorithms based on RGB images and human action recognition algorithms based on skeleton information.In the early stage,RGB images of human movement were obtained by ordinary cameras,and then feature extraction was carried out to obtain action categories.However,this method had poor robustness,low recognition rate,and was easily affected by dynamic background and ambient light.With the proposed key point detection algorithm of RGB images and the advent of depth camera,people can accurately extract the skeleton information of human movement.Skeleton information has the advantages of small data volume,strong generalization ability,and not easy to be affected by the external environment.Using skeleton information as the data source of action recognition can effectively improve the recognition rate of human action.Therefore,the research of this paper is based on the three-dimensional skeleton information.(1)Considering that the current human action recognition algorithm based on RGB images is relatively low and the ST-GCN ignores the different degrees of influence of each point on the action in human action recognition,a two-flow human action recognition algorithm combining convolutional neural network and graph convolutional neural network is proposed.Human action recognition is carried out by extracting the space and time features of human skeleton information.Firstly,the spatial and temporal graph of human skeleton information was constructed,and the improved ST-GCN was used to extract the temporal and spatial characteristics of skeleton information.Secondly,the skeleton information motion graph is constructed,and the features extracted from the convolutional neural network are used as the supplement of the temporal and spatial features of the features extracted from the spatio-temporal graph convolutional network.Finally,the two networks are fused to form a human action recognition algorithm based on the two-flow attention mechanism.The algorithm enhanced the characterization ability of skeleton information and effectively improved the recognition accuracy of human movements.(2)In view of the problems such as fixed topological structure and no use of the length and direction information of bones in the process of the graph convolution construction of ST-GCN,an adaptive mechanism is introduced,and a human action recognition algorithm based on the adaptive mechanism is proposed,and second-order data of bones are added to supplement the information of the original joint data.Finally,a multi-stream human action recognition method based on adaptive mechanism was formed by fusion of skeleton information motion graph to further improve the performance of the algorithm.Finally,fusion experiments were conducted on NTU-RGB+D60 data sets.The recognition rates of(Cross Subject,CS)and(Cross View,CV)based on attention mechanism were 86.5% and 93.5%.Respectively,the recognition rates of the multi-stream human action recognition method based on the adaptive mechanism in CS and CV are 86.9% and 95.0%.Compared with other method,the two action recognition methods proposed in this paper have been improved to some extent.