A CNN-Transformer Hybrid Network for Hand Gesture Classification based on High-Density sEMG

Abstract

In recent years, rehabilitation robots have been gradually applied to assist patients with movement disorders to improve their motor functions. Active training, as an important rehabilitation strategy, can stimulate users to participate in the rehabilitation process more actively and help promote the recovery of motor function. Accurate recognition of motion intention is essential for effective active training. Among various physiological signals, surface electromyography (sEMG) is considered as an ideal signal for gesture intention recognition due to its precision and high real-time performance. sEMG signals can be categorized into two types: sparse sEMG and high-density sEMG (HD-sEMG). Compared with the sparse sEMG, HD-sEMG contains substantial spatio-temporal features and has gradually become an ideal choice for complex multi-gesture classification tasks. However, the complexity of HD-sEMG data also poses challenges to classification methods. Different from machine learning methods based on feature extraction, this paper develops a novel hybrid gesture recognition framework based on CNN-Transformer, which combines the features of convolution and attention mechanisms to enhance network performance. In addition, without complex preprocessing methods such as frequency domain analysis, only data windowing and simple filter are used in the method. The proposed method has been extensively evaluated on a public dataset containing 65 gestures. The results demonstrate substantial improvements over the existing work, with an accuracy of 97.75%.