Electromyogram (EMG) based fingers movement recognition using sparse filtering of wavelet packet coefficients

TLDR: This work proposes an accurate and novel scheme for feature set extraction and projection based on Sparse Filtering of wavelet packet coefficients which can classify a large number of single and multiple finger movements accurately with reduced hardware complexity.

Abstract: Surface electromyogram (EMG) signals collected from amputee’s residual limb have been utilized to control the prosthetic limb movements for many years. The extensive research has been carried out to classify arm and hand movements by many researchers. However, for control of the more dexterous prosthetic hand, controlling of single and multiple prosthetic fingers needs to be focused. The classification of single and multiple finger movements is challenging as the large number of EMG electrodes/channels are required to classify more number of movement classes. Also the misclassification rate increases significantly with the increased number of finger movements. To enable such control, the most informative and discriminative feature set which can accurately differentiate between different finger movements must be extracted. This work proposes an accurate and novel scheme for feature set extraction and projection based on Sparse Filtering of wavelet packet coefficients. Unlike the existing feature extraction-projection techniques, the proposed method can classify a large number of single and multiple finger movements accurately with reduced hardware complexity. The proposed method is compared to other combinations of feature extraction-reduction methods and validated on EMG dataset collected from eight subjects performing 15 different finger movements. The experimental results show the importance of the proposed scheme in comparison with existing feature extraction-projection schemes with an average accuracy of 99.52% ± 0.6%. The results also indicate that the subset of five EMG channels delivers similar accuracy (>99%) to those obtained from all eight channels. The resultant accuracy values are improved over the existing one reported in the literature, whereas only one-third numbers of channels per identified motions are employed. The experimental results and analysis of variance tests (p < 0.001) prove the feasibility of the proposed work.