Action recognition and correction by using EMG signal for health sports

Abstract

When a person perform daily sports or exercises, his or her actions may not be standard. If the actions are non-standard, the effect of sports or exercises would degrade seriously. Thus, recognizing non-standard actions is critical to provide constructive suggestions for daily exercises. In order to recognize actions and correct the non-standard ones during exercises and sports, we propose a method based on EMG signal to provide exercise suggestions. We first divide each channel of EMG signal into fixed-size segments in the form of sliding windows, second use short-time Fourier Transform to convert EMG signal segments as spectrograms, third input the spectrograms into a convolutional neural network to recognize the actions, fourth use a decision tree to determine whether the action is standard or non-standard and provide exercise suggestions if the action is non-standard. The experimental results show that the proposed method could identify most of the actions and recognize the non-standard action.