3D Skeleton-Based Video Action Recognition by Graph Convolution Network
Xuesong Gao, Keqiu Li, Yu Zhang, Qiguang Miao, Lijie Sheng, Jun Feng Xie, Jinfu Xu · 2019
Human action recognition based on 3D skeleton data attracts more and more researchers because of its good robustness on the influence of light and occlusion. In previous studies, people only identified individual skeletons but ignored the fact that an action was usually performed by two or more people. Based on this idea, we trained a single and double person motion model based on 3D skeleton data and achieved a better recognition effect by analyzing the number of people involved in an action in the real-time recognition process. In response to these challenges, we propose a novel end-to-end model based on Graph Convolutional Network (GCN). Our model divides human skeleton into five regions (four limbs and trunk) to extract internal and inter-regional features. Experiments show that our model has reached the current advanced level on the NTU RGB-D dataset.