Research On Human Behavior Recognition Method Based On Graph Convolutional Networks

In recent years,with the rapid development of artificial intelligence,human behavior recognition technology has also attracted people’s attention.How to quickly and accurately recognition human behavior has become one of the important research topics,but many recognition methods of existence has many problems,for example,the environmental impact is large,and the characteristics of human body structure and human movement cannot be effectively used,and the information in the dataset is incomplete.After analyzing the existing research methods and existing problems,this thesis proposes two new methods of human behavior recognition.First of all,in the behavior recognition,aiming at the problem that the accuracy is not high due to the two characteristics of not taking full advantage of the specificity of human body structure and the long-term dependence of action sequences,a method of combining spatio-temporal graph convolutional neural network and long short-term memory network are proposed.Among them,the graph convolutional neural network makes full use of the characteristics of the human body structure,treats the skeleton as a topological graph,and performs convolution from two aspects of time and space to extract features.The long short term memory network can deal with long-term dependence well.The model based on this method is used to process the action sequence and use the connection between frames to improve the recognition accuracy.Secondly,when using skeleton data for human behavior recognition,only the threedimensional coordinates of the joint points are used,and the obtained information is insufficient.In view of this situation,select the appropriate joint points from the human skeleton as the base point,establish the spatial relationship between all joint in the body and the base point,process the two features of the joint points and the spatial relationship,and establish the convolutional neural network of the dual-flow spatio-temporal graph The network will fuse the two classification results to conduct behavior recognition.Finally,the established model is used to recognition behaviors in the NTU RGB + D120 dataset,compared with other experimental methods,the two human behavior recognition methods proposed in this thesis effectively improve the accuracy.