A lightweight graph convolutional network for skeleton-based action recognition

Human action recognition has been an attractive research topic in recent years due to its wide range of applications. Among existing methods, the Graph Convolutional Network achieves remarkable results by exploring the graph nature of skeleton data in both spatial and temporal domains. Noise from the pose estimation error is an inherent issue that could seriously degrade action recognition performance. Existing graph-based methods mainly focus on improving recognition accuracy, whereas low-complexity models are required for application development on devices with limited computation capacity. In this paper, a lightweight model is proposed by pruning layers, adding Feature Fusion and Preset Joint Subset Selection modules. The proposed model takes advantages of the recent Graph-based convolution networks (GCN) and selecting informative joints. Two graph topologies are defined for the selected joints. Extensive experiments are implemented on public datasets to evaluate the performance of the proposed method. Experimental results show that the method outperforms the baselines on the datasets with serious noise in skeleton data. In contrast, the number of parameters in the proposed method is 5.6 times less than the baseline. The proposed lightweight models therefore offer feasible solutions for developing practical applications.