A Wavelet-Based Surface Electromyogram Feature Extraction for Hand Gesture Recognition

Sammanfattning: The research field of robotic prosthetic hands have expanded immensely in the last couple of decades and prostheses are in more commercial use than ever. Classification of hand gestures using sensory data from electromyographic signals in the forearm are primary for any advanced prosthetic hand. Improving classification accuracy could lead to more user friendly and more naturally controlled prostheses. In this thesis, features were extracted from wavelet transform coefficients of four channel electromyographic data and used for classifying ten different hand gestures. Extensive search for suitable combinations of wavelet transform, feature extraction, feature reduction, and classifier was performed and an in-depth comparison between classification results of selected groups of combinations was conducted. Classification results of combinations were carefully evaluated with extensive statistical analysis. It was shown in this study that logarithmic features outperforms non-logarithmic features in terms of classification accuracy. Then a subset of all combinations containing only suitable combinations based on the statistical analysis is presented and the novelty of these results can direct future work for hand gesture recognition in a promising direction.