A Separable Spatial–Temporal Graph Learning Approach for Skeleton-Based Action Recognition

Abstract

With the popularization of sensors and the development of pose estimation algorithms, a skeleton-based action recognition task has gradually become mainstream in human action recognition tasks. The key to solving skeleton-based action recognition task is to extract feature representations that can accurately outline the characteristics of human actions from sensor data. In this letter, we propose a separable spatial-temporal graph learning approach, which is composed of independent spatial and temporal graph networks. In the spatial graph network, spectral-based graph convolutional network is selected to mine spatial features of each moment. In the temporal graph network, a global-local attention mechanism is embedded to excavate interdependence at different times. Extensive experiments are carried out on the NTU-RGB+D and NTU-RGB+D 120 datasets, and the results show that our proposed method outperforms several other baselines.