STCN-GR: Spatial-Temporal Convolutional Networks for Surface-Electromyography-Based Gesture Recognition

Gesture recognition using surface electromyography (sEMG) is the technical core of muscle-computer interface (MCI) in human-computer interaction (HCI), which aims to classify gestures according to signals obtained from human hands. Since sEMG signals are characterized by spatial relevancy and temporal nonstationarity, sEMG-based gesture recognition is a challenging task. Previous works attempt to model this structured information and extract spatial and temporal features, but the results are not satisfactory. To tackle this problem, we proposed spatial-temporal convolutional networks for sEMG-based gesture recognition (STCN-GR). In this paper, the concept of the sEMG graph is first proposed by us to represent sEMG data instead of image and vector sequence adopted by previous works, which provides a new perspective for the research of sEMG-based tasks, not just gesture recognition. Graph convolutional networks (GCNs) and temporal convolutional networks (TCNs) are used in STCN-GR to capture spatial-temporal information. Additionally, the connectivity of the graph can be adjusted adaptively in different layers of networks, which increases the flexibility of networks compared with the fixed graph structure used by original GCNs. On two high-density sEMG (HD-sEMG) datasets and a sparse armband dataset, STCN-GR outperforms previous works and achieves the state-of-the-art, which shows superior performance and powerful generalization ability.