Efficient Surface EMG Compression for Real-time Application

Abstract

Recently, with the development of telecommunication technology, real-time applications using surface electromyogram (sEMG) such as telemonitoring, sEMG bio-feedback and active prosthetics have been developed in many field. However, large sEMG dataset that are transmitted and processed cause a lot of power consumption in real-time applications. In order to reduce the data size of sEMG, I propose an effective compression process. First, the sEMG signal is under-sampled at 500 Hz and a compression algorithm is applied to the sampled signal. There are three compression algorithms used: discrete wavelet transform (DWT), compressive sensing (CS), and compressive covariance sensing (CCS). The proposed process is verified by gesture classification of sEMG datasets of 13 volunteer subjects. Each dataset consists of 6 wrist gestures and 4 finger gestures. In addition, non-linear support vector machine (SVM) with three features extracted from reconstructed signal is utilized for gesture classification. Using three compression algorithms, the sEMG data achieved a compression ratio (CR) in the range of 0.5-0.15. In all gesture classification, DWT showed the highest classification accuracy of more than 85% at the 1/3 CR. Furthermore, DWT’s classification accuracy of wrist gestures was higher than using raw sEMG signal. However, classification accuracy was reduced in finger gesture classification which has small and delicate sEMG signal. CCS is not affected by the amplitude of signal because it reconstructs the covariance of sEMG. Thus, there is no reduction of accuracy in finger gesture classification. As a conclusion of this study, DWT and CS show good performance for clean and large signals, but CCS is a good alternative for small and delicate signals.