sEMG based segmentation parameter influence on Hand gesture recognition using Deep Learning
Hand gesture recognition based on surface electromyography (sEMG) is frequently utilised in artificial prostheses, rehabilitation training, and human-computer interfaces. Although deep learning based classification of sEMG has yielded fairly acceptable outcomes, the process of sEMG signal segmentation is typically led by heuristics, and is an under-investigated problem with implications on optimal data size, model selection and real-time applications. Initially, we developed a 1D CNN model that distinguishes seven hand motions from multi-channel sEMG obtained using forearm positioned myo-sensor. We then present a detailed analysis of various segmentation parameters and how they affect the accuracy of categorizing hand gestures. The observed F1-scores of the model highlights that smaller window size of 200 ms provides a better classification performance compared to larger window sizes, with possible performance stagnation beyond 1000 to 2000 ms. This finding potentially highlights that muscle activation for each gesture carry the imprint of that gesture, even early in the action, and hence not requiring large windows for final classification while using deep learning techniques.
Future Works
Development of a deep learning system to identify the completion status of arm rehabilitation exercises for erb’s palsy patients.
