Robust upper limb motion classification using Gaussian mixture models

Abstract

A Gaussian mixture model (GMM) based classification scheme is proposed in this paper to perform multiple limb motion discrimination using continuous myoelectric signals (MES) from limb muscles. The system is optimized with respect to the feature set, classifier and post-end processing of the decisions through comprehensive experimentation. The experiments examine the effects of various feature sets including the time-domain (TD) features and the autoregressive (AR) features with root mean square value (RMS), and the effect of the majority vote (MV) in post-processing on the classification performance. The averaged GMM classification performance is compared with that of three other motion techniques (a linear discriminant analysis (LDA), a linear perceptron (LP) neural network and a multilayer perceptron (MLP) neural network). The Gaussian mixture motion model achieves 96.91% classification accuracy using a combination of AR with RMS and TD (AR+RMS+TD) feature set for a six class problem. It has been demonstrated that this GMMbased limb motion classification scheme has superior classification accuracy and results in a robust method of motion classification.