IMU Data-based Recognition for Sports Exercises: An Enhanced Distance Optimization Approach for Repetition Counting ​Across Activities

Abstract

In the field of human activity recognition (HAR), inertial measurement units (IMUs) are a commonly used method to record movement patterns. In the study presented in this paper, IMU captured seven different sports exercises performed by 21 participants. In the preliminary data analysis phase, an exercise classification was conducted using the Long Short Term Memory (LSTM) Network and Temporal Convolutional Network (TCN). The LSTM achieved an accuracy rate of 94.2 % for training and 90.8 % for testing. Similarly, the TCN demonstrated rates of 95.5 % for training and 91.6 % for testing. The subsequent stage was centered on quantifying the number of completed repetitions. A distance value was derived which showed promising results for exercise-independent counting without the need for manual feature selection. For further improvement, a range-to-mean ratio of the standard deviation was calculated and used for feature selection. Combined with a local extrema analysis of the modified distance values the accuracy of counting repetitions was significantly improved, especially for exercises that show irregularities in the signal course.

Keywords

• human activity recognition
• inertial sensors
• artificial neural networks
• repetition counting
• distance
• local extrema analysis