Robust In-the-Wild Exercise Recognition from a Single Wearable: Data-Side Fusion, Sensor Rotation, and Feature Engineering

TL;DR

This paper presents a robust single-sensor exercise recognition method that employs feature extraction, data augmentation, and sensor data fusion to improve accuracy in unconstrained, real-world conditions.

Contribution

It introduces a novel data-side fusion and augmentation approach combined with a soft voting classifier for improved exercise recognition from a single wearable sensor.

Findings

Achieved 58.83% macro F1 score overall

Fractal and spectral features are most important for arm recognition

Sensor rotation and axis inversion improve model robustness

Abstract

Monitoring physical exercises is vital for health promotion, with automated systems becoming standard in personal health surveillance. However, sensor placement variability and unconstrained movements limit their effectiveness. This study proposes the team "3KA"'s one-sensor workout activity recognition method using feature extraction and data augmentation in 2ndWEAR Dataset Challenge. From raw acceleration, angle and signal magnitude vector features were derived, followed by extraction of statistical, fractal/spectral, and higher-order differential features. A fused dataset combining left/right limb data was created, and augmented via sensor rotation and axis inversion. We utilized a soft voting model combining Hist Gradient Boosting with balanced weights and Extreme Gradient Boosting without. Under group 5-fold evaluation, the model achieved 58.83\% macro F1 overall (61.72% arm, 55.95% leg). ANOVA F-score showed fractal/spectral features were most important for arm-based recognition but least for leg-based. The code to reproduce the experiments is publicly available via: https://github.com/Khanghcmut/WEAR\_3K