Time domain multi-feature extraction and classification of human hand movements using surface EMG

Hand movement recognition from electromyogram (EMG) signals is a crucial element for design of electrically controlled limb prostheses and for developing human computer interfaces (HCIs). The study focuses on extraction of significant features from raw EMG signals corresponding to six different grasping hand movements and then using them to classify the respective hand movements using different classifiers. We propose a hybrid multi-feature set comprising of Autoregressive (AR), Root Mean Square (RMS), Zero Crossing (ZC), Slope Sign Change (SSC), Waveform Length (WL) and Mean Absolute Value (MAV) time domain features. We use four different classifiers (k-NN, LDA, QDA and Subspace Discriminant Ensemble) for this experiment. Different set of features yield varying accuracies depending upon the choice of classifiers. Conventional AR feature set provides maximum accuracy of 80.83% with LDA classifier. Similarly, the feature set excluding AR obtains maximum accuracy of 72.5% with k-NN classifier. Our proposed multi-feature set of all six features provides the highest accuracy of 83.33% with Ensemble classifier which is significantly higher than the accuracy values of other feature sets. We validate the results on a large dataset of 600 sEMG signals corresponding to 6 different grasping hand movements.