Free Weight Exercises Recognition Based on Dynamic Time Warping of Acceleration Data

Abstract

To maximize training effects in free weight exercises, people need to remember repetitions of each type of exercises, which is tedious and difficult. Recognizing exercises type and counting automatically can overcome this problem, and multiple accelerometers were used in the existing exercises recognition. This paper presents a new recognition method based on one tri-axial accelerometer, in which a filtered acceleration data stream is divided into time series with unequal length for peak analysis instead of conventional fixed length window. Based on this time series, Dynamic Time Warping (DTW) is deployed to recognize weight exercise types. 3D Euclidean distance and Itakura parallelogram constraint region are used to improve recognition performance. A reference template is set up for each class based on many examples instead of one in the conventional way. The proposed procedures are compared with other popular methods with both the user-dependent protocol and the user-independent protocol. Results show that proposed approach is feasible and can achieve good performance.