Research On Human Skeleton Behavior Recognition Method Based On Graph Convolutional Network

Human behavior recognition technology is a popular research direction in the field of computer vision.With the rapid development of artificial intelligence and the Internet,this technology has a wide application prospect and economic value,and is widely used in intelligent monitoring systems,virtual reality,human-computer interaction and other fields.The core of human behavior recognition technology is the establishment of behavior recognition data set and the extraction of behavior features.With the development of human skeletal joint point extraction technology and graph convolutional neural network(GCN),GCN-based human 3D skeletal behavior recognition technology has become a new research direction.The model maps used in the currently existing graph convolution-based human behavior recognition methods have certain localization and are not ideal for the extraction of skeletal sequence features,which leads to the lack of global modeling capability of these methods,difficulty in adapting to the diversity of data in action recognition tasks,and poor performance in terms of model recognition accuracy.To address the above problems,the main research work of this thesis is as follows:(1)A spatio-temporal dimensional adaptive graph convolutional network model(STAGCN)is proposed.ST-AGCN is an improved model for most models using a fixed graph topology.The model generates variable graphs by training and updating the fixed graph in different layers in both spatial and temporal dimensions.The nodes and edges in the variable graph are adaptively adjusted to establish corresponding connections based on the dependencies of remote joints.Subsequently,the model is experimented on NTU-RGB+D dataset for skeleton behavior recognition,which demonstrates its excellent recognition accuracy and good generalization performance.(2)A graph-adaptive spatio-temporal attention graph convolutional model(STA-AGCN)is proposed.The model is based on the adaptive graph convolutional network model with the integration of the spatio-temporal attention module.The spatio-temporal attention module utilizes the principle of continuity during human motion to generate attention masks in space using the feature differences between skeletal points of the preceding and following frames,and extracts the spatial features of each frame in the temporal domain to generate an attention tensor for filtering key frames while better fusing spatial and temporal information.Through experiments,the model is compared on open source datasets NTU-RGBD and Kinetics,and it is found that the proposed STA-AGCN model has a good improvement in recognition accuracy and a great reduction in computational effort,which confirms that the proposed model is effective and feasible.