Classification of sEMG signals of hand gestures based on energy features

The performance of a robotic exoskeleton depends upon the accuracy of control commands from the controller fed with Surface ElectroMyoGraphy (sEMG) input signals. The classification of hand gestures based on sEMG signals extracted from a human hand depends upon the type of EMG features extracted. In this paper, an ensemble of energy features is proposed for the sEMG classification. The idea is motivated by the energy features’ relation to the movement force, dependence on related mechanical factors, robustness with respect to the repetition of trials and the presence of noise. The suitability of the proposed energy features is tested by using the standard machine learning classifiers, including the K-Nearest Neighbour (KNN), Probabilistic Neural Networks, Ensemble KNN, Quadratic Discriminant Analysis and the Cubic Support Vector Machines. In order to show the superiority of the proposed energy features, the experiments are conducted over benchmark NinaPro DB1 sEMG hand gesture dataset. The fine KNN classifier has achieved the highest validation accuracy of 88.8%, an improvement of 13% over the state of the art accuracy. The performance of the classifiers is analyzed with various evaluation metrics using the proposed feature ensemble. The contribution of individual features for the performance is also analyzed and observed that spectral band energy features have provided an highest accuracy of 85.2%. Additionally, the proposed method is found to be computationally least expensive.