An Efficient Graph Convolution Network for Skeleton-Based Dynamic Hand Gesture Recognition

Abstract

Dynamic hand gesture recognition has evolved as a prominent topic of computer vision research due to its vast applications in human–computer interaction, robotics, and other domains. Although there are numerous related recognition studies, the state-of-the-art (SOTA) methods are over-parametrized. Specifically, the number of model parameters is quite large, which results in high-computational costs. This work, referring to Song’s ResGCN, designs an efficient and lightweight graph convolutional network (GCN), named ResGCNeXt. ResGCNeXt learns rich features from skeleton information and achieves high accuracy with less number of model parameters. First, three data preprocessing strategies according to motion analysis are designed to provide sufficient features for the recognition model. Then, an efficient GCN structure combining bottleneck and group convolution is designed to reduce the number of model parameters without loss of accuracy. Furthermore, an attention block called SENet-part attention (SEPA) is added to improve channel and spatial feature learning. This study is validated on two benchmark data sets, and the experimental results show that ResGCNeXt provides competitive performance, especially, in significantly reducing the number of model parameters. Compared to HAN-2S, which is one of the best SOTA methods, our method has half model parameters and a 0.3% higher recognition rate.