Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Hoang Khang Phan, Khang Le, Tu Nhat Khang Nguyen, Anh Van Dao, Nhat Tan Le

TL;DR

This work tackles robust one-sensor exercise recognition under real-world sensor-placement variability by fusing contralateral limbs and applying axis inversion and sensor-rotation augmentations. It builds a large feature set from raw accelerometer data, signal magnitude vector representations, and angle-based measures, then uses a soft voting ensemble of HGBC and XGBoost, evaluated with group-5 cross-validation on the WEAR dataset. The approach yields macro-F1 around 58.8% overall (61.7% for arm, 55.95% for leg), with fractal/spectral features especially informative for arm movements and cross-limb fusion helping distinguish similar activities. The study highlights practical robustness in the wild, discusses current limitations, and proposes gyroscope integration and generative data synthesis as promising future directions; code is publicly available for reproduction.

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

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

TL;DR

This work tackles robust one-sensor exercise recognition under real-world sensor-placement variability by fusing contralateral limbs and applying axis inversion and sensor-rotation augmentations. It builds a large feature set from raw accelerometer data, signal magnitude vector representations, and angle-based measures, then uses a soft voting ensemble of HGBC and XGBoost, evaluated with group-5 cross-validation on the WEAR dataset. The approach yields macro-F1 around 58.8% overall (61.7% for arm, 55.95% for leg), with fractal/spectral features especially informative for arm movements and cross-limb fusion helping distinguish similar activities. The study highlights practical robustness in the wild, discusses current limitations, and proposes gyroscope integration and generative data synthesis as promising future directions; code is publicly available for reproduction.

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

Paper Structure

This paper contains 16 sections, 3 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. Data Augmentation Strategy
  5. Feature extraction
  6. Raw-accelerometer-based features
  7. Signal-Magnitude-Vector-based features
  8. Angle-based features
  9. Feature Normalization
  10. Classification Model
  11. Evaluation settings
  12. Results
  13. Classification performance
  14. Feature analysis
  15. Discussion
  16. ...and 1 more sections

Figures (4)

  • Figure 1: The flowchart of the proposed pipeline. For each limb, data from the right and left sides were first combined. Data augmentation was then applied to the merged data, followed by feature engineering. Finally, a single model was trained per limb, making the pipeline practical for real-world use.
  • Figure 2: The confusion matrix of activity recognition by (a) arm and (b) leg.
  • Figure 3: The box plot of ANOVA F score by group of feature
  • Figure 4: The camera footage of (a) subject 15 and (b) subject 18 during their abnormal position triceps stretching activity session.