Hand Gesture Recognition by a MMG-Based Wearable Device

Abstract

A novel wearable human machine interface based on mechanomyogram (MMG) signals was presented in this study. A three-axis accelerometer was fixed to a customized watch strap to measure the MMG signals that were generated by the end of the extensor digitorum muscle. Eight gaming gestures, including clapping, index figure flicking, finger snapping, coin flipping, shooting, wrist extension, wrist flexion and fist-making, were identified in real time. This study extracted the features from both the time signals and the coefficients of the wavelet packet decomposition (WPD), and sequential forward selection (SFS) was used to identify the significant features to improve the classification accuracy and reduce the processing time. The performances of the classifiers such as the k-nearest neighbors (KNN), the support vector machine (SVM), linear discriminant analysis (LDA), and deep neural network (DNN) were compared. After testing the system on 35 subjects aged from 16 to 55 years old, the proposed system has advantages with respect to its convenient portability, stable signal acquisition, low power consumption, and high classification accuracy.