Reduce sEMG channels for Hand Gesture Recognition

Multi-channel surface electromyography (sEMG) acquisition devices are generally used for hand gesture recognition in the area of human-computer interaction, rehabilitation training, and artificial prosthesis. Multi-channel sEMG can capture abundant information related to muscle motion but at the cost of increased complexity and signal crosstalk. In this work, we aim to use sEMG of fewer channels to realize recognition accuracy comparable to that with more channels. Specifically, time-domain features extracted from sEMG of multiple channels are first evaluated by Sequential Forward Feature Selection method based on Mutual Information (SFFSMI), where we can achieve gesture recognition rate of 99.64%. We then apply a channel selection method by combining Relief-F algorithm with support vector machine classifier (SVM-MRCS for short) and get fewer channels. In order to maintain recognition performance comparable to that with more channels, the wavelet and wavelet packet are further used to extract features fed to a SVM classifier. Experimental results show that we can use only four out of eight channels to obtain gesture recognition accuracy of 99.53%, which well balances the recognition performance and sEMG device complexity.