Machine learning for IMU motion tracking

Study overview

Inertial measurement units, or IMUs, are widely used to track human movement in rehabilitation, biomechanics, activity monitoring, and healthcare applications. They are portable and can support real-time feedback, but their accuracy can decline over time because small sensor errors accumulate during motion tracking.  

This study evaluated whether machine learning could improve IMU-based position tracking for human movement. The researchers tested machine learning models for two related tasks: identifying moments when the sensors were stopped, known as zero-velocity detection, and estimating sensor displacement during movement.  

The study compared machine learning approaches, including Random Forest, Support Vector Machine, and Long Short-Term Memory models, with traditional threshold-based and drift-correction methods. The goal was to understand where machine learning could improve motion tracking accuracy and where conventional approaches remained competitive.

Key findings

Machine learning improved zero-velocity detection

Machine learning classifiers were better than traditional fixed-threshold methods at identifying stationary periods in IMU data.

This matters because detecting when a sensor is not moving can help reduce drift and improve motion tracking accuracy over time.  

Motion estimation remained challenging

LSTM-based regression models were tested to estimate sensor displacement during movement, but they did not significantly outperform traditional drift-corrected double integration.

This finding is important because it shows where machine learning helped and where more established motion-tracking methods remained competitive.  

A combined neural network showed promise

The study also tested a stacked neural network that combined zero-velocity detection and motion estimation. This combined model showed lower average position-tracking error than the separate-method approaches.  

Vertical motion was easier to track than horizontal motion

The study found that vertical movement was tracked more accurately than horizontal movement, likely because horizontal tracking is more affected by accumulated errors over time.  

Why this study matters

This study reflects Sword’s early work in making motion tracking more accurate, scalable, and useful for care delivered outside traditional clinical settings.

Accurate movement tracking is foundational to digital rehabilitation. It helps clinicians and care teams understand how patients are moving, whether exercises are being performed correctly, and how progress changes over time. For technology-enabled care to be clinically meaningful, motion data needs to be reliable enough to support feedback, monitoring, and decision-making.

The study found that machine learning can improve some parts of IMU-based motion tracking, especially zero-velocity detection, while also showing that not every motion-tracking challenge is solved by machine learning alone. That balance is important. It positions AI and machine learning as tools that can strengthen clinical technology when used thoughtfully, rather than as a blanket replacement for validated methods.

For Sword’s clinical research library, this study belongs in the AI care category because it explores how machine learning can improve the technical foundation of sensor-guided rehabilitation.

Read the full study