This study investigates how IMU sensor data types, feature extraction, window sizes, and supervised machine learning models affect swimming activity recognition accuracy. Using a single IMU on lower back, the system classified strokes, turns, and starts under realistic conditions. Three classifiers of Support Vector Machine (SVM), Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA) were evaluated. Five types of statistical feature sets were extracted from window sizes ranging from 1 to 6 seconds with 0.01 second sliding step. The best classifier is SVM, achieving overall accuracy of 0.97 with high accuracy for surface swimming stroke. However, shorter windows had better accuracy for activity of start. The approach shows potential for real-time swimmer performance analysis and broader sports activity recognition.
© Copyright 2025 ISBS Proceedings Archive: Vol. 43: Iss. 1. All rights reserved.