BACKGROUND: Automated swimming activity classification using inertial measurement units (IMU) enables real-time session monitoring. However, classification may depend on sensor placement and model training/testing strategy. Therefore, this study aims to examine the impact of IMU sensor placement (head and lower-back) and two training/testing strategies: a subject-independent approach using Leave-One-Subject-Out (LOSO) cross-validation, where each swimmer is evaluated using a model trained on the remaining swimmers, and a subject-dependent strategy, where each swimmer`s Individual Medley data served as the test set; METHODS: A Support Vector Machine classifier was trained using mean and standard deviation features extracted from 0.3-second windows at 100 Hz and applied to nine activity classes across 12 swimmers. Per-swimmer accuracy and per-class precision, recall, and F1-scores are reported across four experimental conditions; RESULTS: A consistent performance hierarchy was observed: lower-back sensor with Individual Medley testing produced the highest performance, followed by head sensor with Individual Medley testing, lower-back sensor with LOSO, and head sensor with LOSO. The head sensor with LOSO showed the greatest inter-swimmer accuracy variability (0.43-0.9). Among all classes rest and stroke styles were classified most accurately with Individual Medley testing; CONCLUSIONS: The Individual Medley testing outperformed LOSO by 4% on lower-back and 10% on head sensor. Lower-back sensor with Individual Medley testing achieved high performance for frontcrawl, backstroke and butterfly while the head sensor with Individual Medley testing strategy reached comparable performance. Wall push-off and turn remained challenging to classify with all methods that indicate a limitation in capturing short activities.
© Copyright 2026 XVth International Symposium on Biomechanics and Medicine in Swimming Proceedings. Alle Rechte vorbehalten.