Fast distance-enhanced graph convolutional network for skeleton-based action recognition

Exploring how to use graph structures more effectively for extracting and expressing skeleton information has become a significant area of skeleton-based human action recognition. Existing research mostly focuses on adjacent or local skeletal points, which might limit the receptive field (sampling range of skeleton point influence) of graph convolutional networks. In this paper, we introduce a fast distance-enhanced graph convolutional network (FD-GCN) by expanding the impact of remote skeleton joints for relative skeletons’ topological graph. FD-GCN expresses the skeleton information of the topological graph better by reasonably expanding the receptive field. In addition, skeleton-based action recognition becomes increasingly redundant after adding various modules and runs very slowly. Therefore, FD-GCN introduces a temporal convolution structure that can assign position information and weights to time frames to solve this problem. FD-GCN achieves excellent results on NTU RGB+D, NTU RGB+D 120 and NW-UCLA datasets. FD-GCN matches the SOTA performance on the X-Sub benchmark of NTU120, achieving 89.4% accuracy. In addition, FD-GCN achieves 97.0% accuracy on the NW-UCLA dataset, outperforming the previous SOTA by 0.2% in a 4-stream setting. More importantly, the inference time of FD-GCN is about 1.6 ms, while the inference time of the SOTA methods are mostly more than 10 ms.