Hand Gesture Recognition From Surface Electromyography Signals With Graph Convolutional Network and Attention Mechanisms

In human body action recognition, graph convolutional networks (GCNs) show remarkable capability, compared to other deep learning (DL) methods, in capturing latent correlations within the human body topology. However, GCN methods have been rarely studied for hand gesture recognition (HGR) using surface electromyography (sEMG) because it is challenging to define a reliable topology across implicit muscle networks. In this article, we propose a novel covariance-based topology refinement module (CovTRM) to enable the GCN model to adaptively learn the dynamic topologies for various hand gestures. Extensive evaluations of two datasets, the Ninapro DB2 dataset and the UOW Dataset, show that the CovTRM can effectively refine the topologies to adapt to the implicit muscle synergies of different hand gestures. The proposed covariance-based graph convolutional network (CovGCN) model outperforms many machine learning (ML) models in recognizing sEMG-based hand gestures and mitigating the impact of variable limb positions, thereby contributing toward more effective and adaptable prosthetic hand control systems.