Research On Algoriehm Of Human Action Recognition Based On Spatial Temporal Graph Convolution

With the research and rapid development of artificial intelligence technology,human action recognition develops gradually.Human action contains a wealth of behavioral information and is an important source of information for human-computer interaction.How to judge human behavior quickly and accurately is also an important problem in current research.In this paper,the action recognition based on spatialtemporal graph convolution was studied,and a human-computer interaction system with face recognition was designed by combining with Nao robot in the laboratory.Firstly,the YOLOv4 algorithm was studied,and a target detection framework was built based on the YOLOv4-Tiny algorithm.On the basis of this framework,SPP algorithm and res-net structure were integrated.so as to improve the detection accuracy while ensuring the detection speed.The algorithm was combined with the Deepsort target tracking algorithm for realizing target recognition and tracking.After that,the human pose estimation algorithm was studied,the performance of Alphapose algorithm and Open Pose algorithm were compared.The high efficiency of the Alphapose algorithm was verified by experiments.Then,the human action recognition algorithm based on spatial temporal graph convolution and the improved double-flow spatial temporal graph convolution method were studied.Smart-home data set was used for training.The data set was shot in the real environment,which had a certain authenticity and reflects the human behavior information in daily life.The improved target detection algorithm and pose estimation algorithm were combined with the double-flow graph convolution algorithm,and the feasibility of the proposed algorithm was verified by experiments.In order to meet the practical needs,the LBP algorithm was used for face recognition,and the practical feasibility of the algorithm was verified.Finally,a human behavior simulation experiment was designed by using the existing Nao robot and Alphapose algorithm.Firstly,the position information of each joint of human body was obtained by using binocular ranging method,and then the algorithm was designed to convert it into the joint Angle of Nao robot,which was input into the Nao robot to achieve behavior imitation.