Abstract: Gesture classification based on surface electromyography (sEMG) have been extensively investigated towards the control of smart prosthetic hands. However, most studies did not consider the effects of arm position and movements that frequently occur in daily activities. This study aims to address the gesture classification challenge under arm movements. We collected sEMG and acceleration (ACC) data from fourteen participants, including two individuals with radial artery amputations, while performing gestures in both static and dynamic arm states. Using the collected data, the performance of three machine learning methods was evaluated for gesture classification under both arm movements and static arm conditions. The results revealed a 17.48% decrease in average classification accuracy in the dynamic state compared to the static state when using sEMG signals. Subsequently, the improvement in classification accuracy under arm movements was validated using both sEMG and ACC, with deep learning achieving the highest average accuracy of 84.35% across healthy subjects. Additionally, the study assessed the impact of gesture similarity on classification performance and evaluated the practical efficacy of classifiers for amputees. To further enhance gesture classification accuracy under arm movements, a two-stage gesture classification model training method based on ResNet18 was proposed. This method first learns a generalized motion prototype from population data and then adapts it to individual subjects via fine-tuning, resulting in a 2.49% improvement in average recognition accuracy across all 14 subjects.

Key words: Gesture classification')">Gesture classification, SEMG, ACC, Arm movement, Amputee, Two-stage learning