EMG pattern recognition using decomposition techniques for constructing multiclass classifiers

To improve the dexterity of multi-functional myoelectric prosthetic hand, more accurate hand gesture recognition based on surface electromyographic (sEMG) signal is needed. This paper evaluates two types of time-domain EMG features, one independent feature and one combined feature including four features. The selected features from eight subjects with 13 finger movements were tested with four decomposed multi-class support vector machines (SVM), four decomposed linear discriminant analyses (LDA) and a multi-class LDA. The classification accuracy, training, and classification time are compared. The results have shown that the combined features decrease error rate, and binary tree based decomposition multiclass classifiers yield the highest classification success rate (88.2%) with relatively low training and classification time.