Wearable IMU-based Gym Exercise Recognition Using Data Fusion Methods

Abstract

Gym exercise has become a focus of attention nowadays because of its health benefits. Automatic gym exercise recognition is an emerging research field which aimed at guiding people to keep fit scientifically through gym exercise monitoring. However, as the actions of gym exercise (e.g. barbell bench press, leg extension, etc.) are more diversity and complexity than outdoor exercise (e.g. running, cycling, etc.). Previous studies increase the number of sensors to improve the accuracy gym exercise recognition, while wearing too many sensors make the subjects uncomfortable during gym exercise. In this study, we studied the performance of different classifiers on gym exercise recognition, then the impact of the number of sensors and the positions of the sensor on the body on the recognition performance was analyzed based on the Extra Trees (ET) classifier. Finally, a stratification fusion method using only two sensors was proposed according to the analysis results. The experimental results showed that when two sensors were used to identify eight kinds of gym exercises, the accuracy of the proposed stratification fusion method was 91.26%.