Skeleton-based action recognition based on spatio-temporal adaptive graph convolutional neural-network

To solve the problems of lacking capability to model global context temporal information, insufficient classification accuracy and generalization ability in skeleton-based action recognition, a temporal modeling feature of temporal action graph and a skeleton-based action recognition model based on spatio-temporal adaptive graph convolutional neural-network (ST-AGCN) were proposed. First, graph representation theory and skeleton sequences were introduced, and temporal action graph and adjacent matrix based on N-order fixed-time structure was designed. Then, temporal adaptive graph convolutional network (T-AGCN) structure based on non-local construction combined with graph convolution theory was proposed. Furthermore, through combining T-AGCN with spatial adaptive graph convolutional network (S-AGCN) structure, the skeleton-based action recognition model based on ST-AGCN was proposed. Finally, researches on skeleton-based action recognition were conducted on the NTU-RGB+D and SBU dataset to validate advantages of model in modeling global context temporal information, classification accuracy and generalization ability. Experiment results show that the model could achieve highest accuracy of 92.1% and 99.5% respectively on the above-mentioned two datasets, demonstrating the excellent accuracy and better generalization ability of the model. © 2020, Editorial Board of Journal of Huazhong University of Science and Technology. All right reserved.