Machine Learning Improvements to Human Motion Tracking with IMUs

Study Overview

IMU-based motion tracking is widely used in biomechanics, rehabilitation, and activity monitoring but suffers from error accumulation due to sensor drift. This study explores machine learning (ML) methods to improve accuracy by:

1. Zero-velocity detection – ML classifiers (Random Forest, SVM, LSTM) outperform traditional threshold-based methods in identifying stationary periods, reducing drift.
2. Motion estimation – LSTM-based regression models were tested, but did not significantly outperform traditional drift-corrected double integration.

A stacked neural network combining both tasks showed promising results. Vertical (z-axis) motion was tracked more accurately than horizontal motion due to fewer accumulated errors.

Key Takeaways:

• ML improves zero-velocity detection but struggles with motion estimation.
• Traditional drift correction remains competitive for position tracking.
• Future research should enhance sensor fusion, ground truth accuracy, and model generalization across different movements and IMU placements.

These findings support better real-time motion tracking, benefiting telerehabilitation and healthcare applications.

Read the full study