Artificial neural network classifier in comparison with LDA and LS-SVM classifiers to recognize 52 hand postures and movements

Classification of postures and movements of distal limbs based on surface electromyography (sEMG) of proximal muscles is necessary in myoelectric hand prostheses. With increasing the number of movements, classification problem becomes a serious challenge. In this paper, we have used NINAPRO database that contains sEMG and kinematic data of upper limbs while performing 52 hand postures and movements. We evaluated the performance of MLP classifier in comparison with LDA and LS-SVM classifiers using different combinations of features. First by windowing the signal with two different methods, the major part of the signal was selected and eight various temporal features (MAV, IAV, RMS, WL, E, ER1, ER2, CC) were extracted. Then to achieve the best performance of each classifier, they were evaluated with single, double and multiple combinations of features. For MLP classifier, the best average classification accuracy of 96.34% was achieved for first windowing method and using combination of all features. The corresponding accuracy for LDA classifier (with first windowing method and MAV (or IAV) +CC features) was 84.23%. For LS-SVM classifier (with second windowing method and IAV+MAV+RMS+WL features), the best accuracy of 85.19% was obtained.