Human Action Recognition with Sparse Joints Graph Convolutional Network

Many studies of human action recognition rely on detecting human skeleton to recognize human actions, but when the trajectories of skeleton for different actions are similar, such as drinking and brushing teeth, both involving hand movements towards the head, there is not much distinction, which can lead to confusion and incorrect classification results. This approach aims to address the problem of confounding similar skeletal joints movements that cannot be distinguished. However, in previous human action recognition methods, the skeletal branch was constrained by the human body’s connectivity structure, limiting the extracted features to local regions. This paper introduces a sparse adjacency matrix to allow information to propagate further between joints, enabling the extraction of global features. Experimental comparisons across multiple datasets verify that the proposed approach outperforms previous research in human action recognition.