Two-Steam Fully Connected Graph Convolutional Network for Skeleton-Based Action Recognition

Skeleton-based human action recognition has recently drawn a lot of attentions with the increasing availability of large-scale skeleton datasets. Graph Convolutional Network (GCN) methods have achieved relatively good performances in action recognition. However, most GCN methods based on predefined graphs with fixed topology constraints always neglect the potential dependencies derived from the cooperative movement of all joints. Besides, the lengths and the directions of skeletons are rarely involved. These easily cause a larger deviation of the estimated action from the actual action. Here, a two-steam fully connected graph convolutional network (2s-FGCN) is proposed. The topology structure of the 2s-FGCN covers the local physical connections and the global potential cooperation of all joints and the joints, lengths and directions of skeletons are all input to the model. The experimental results on two datasets (NTU-RGB+D and Kinetics-Skeleton) demonstrate that the proposed model can obtain the state-of-the-art results.