Experimental Considerations on Signal Feature and Kernel/Prameters of SVM in Hand Motion Classification from sEMG

Abstract

This paper introduces the Support Vector Machine (SVM) to classify finger motion patterns from surface EMG (Electromyography). Surface EMG (sEMG) contains several signals from different muscles around the electrode, which make it difficult to estimate actually produced motions. To enhance the classification performance, we investigated which feature of EMG signals is more effective of the following six: raw data, integrated EMG, voltage level difference, power spectrum, FFT peak frequency, or wavelet coefficient. Next, we also considered the selection of SVM's kernel and its parameters. We experimentally demonstrated that the “Voltage level difference” provides 95% or more correct identification rate when the radial basis function (RBF) is utilized as the kernel. Changing the parameter values of the RBF, 98% correct classification rate was obtained in our experiment from three subjects.