An Adaptive Feature Extractor for Gesture SEMG Recognition

This paper proposes an adaptive feature extraction method for pattern recognition of hand gesture action sEMG to enhance the reusability of myoelectric control. The feature extractor is based on wavelet packet transform and Local Discriminant Basis (LDB) algorithms to select several optimized decomposition subspaces of origin SEMG waveforms caused by hand gesture motions. Then the square roots of mean energy of signal in those subspaces are calculated to form the feature vector. In data acquisition experiments, five healthy subjects implement six kinds of hand motions every day for a week. The recognition results of hand gesture on the basis of the measured SEMG signals from different use sessions demonstrate that the feature extractor is effective. Our work is valuable for the realization of myoelectric control system in rehabilitation and other medical applications.