Multi-stage Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition

Skeleton-based action recognition has received a lot of attention because it is independent of the physical features of the human body and it conveys important information about human behavior recognition in simple and clear way. Most of the existing approaches only use spatial or temporal models to extract features. However, that cannot capture the complex regional spatio-temporal joint dependence, which limited its sensory field. To address this problem, a novel model of dynamic skeletons called Multi-Stage Temporal Graph Convolutional Networks (SMS-GCN) is proposed, which consists of the graph convolutional network (GCN) model as well as the multi-stage temporal convolutional network (MS-TCN) model. The MS-TCN model consists of multiple stages of single-stage TCNs and incorporates a bottleneck design to extract temporal information to achieve aggregation of temporal information. The SMS-GCN model achieved an accuracy of 36.7
^%
and 93.13% on the datasets Kinetics and NTU-RGB+D respectively. The experimental results indicate that it achieves substantial improvements over mainstream methods.