Improving sEMG-Based Hand Gesture Recognition Using Maximal Overlap Discrete Wavelet Transform and an Autoencoder Neural Network

In this work we present a preliminary study regarding the use of MODWT decomposition and time domain parameters for the task of hand gesture classification using sEMG signals. A total of 28 features were extracted for each subject and a first test has shown an improvement of up to 5.8% in relation to previously work. As expected, a simple linear classifier (LDA) obtained the best results. As a second test, we attempted to evaluate the classifiers with respect to the prediction of each hand gesture and mixing the individual data of the subjects. Unlike the first test, with a more complex approach, the Autoencoder technique performed an average accuracy of 77.96% +/- 1.24 against only 65.36% +/- 1.09 achieved by the LDA classifier. Each classifier fails mainly to separate the gestures belonging to the same grasp group: precision grasp (Tip, Palmar and Lateral) and power grasp (Cylindrical, Hook and Spherical), because of their similarities from a muscular point of view. This result led us to perform a final test considering only precision/power classification, reaching an average accuracy of 95.60% with the Autoencoder Neural Network. In general, the results illustrate that the method presented in this paper can be applied to real applications; but a more refined approach, with more subjects and gestures, should be done.