Estimation of kinematics from inertial measurement units using a combined deep learning and optimization framework

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Estimating joint kinematics using inertial measurement units (IMUs) is challenging due to magnetometer interference and calibration complexities in uncontrolled environments.
• Traditional sensor-fusion filters require precise sensor alignment and calibration, limiting practical applications in real-world biomechanics.
• Accurate joint angle estimation is crucial for understanding human movement in clinical and sports settings.

Purpose of the Study:

• To develop and validate a novel framework for accurate lower extremity joint angle prediction directly from inertial sensor data.
• To overcome limitations of traditional methods by eliminating reliance on magnetometer readings and simplifying calibration.
• To enable robust and convenient gait kinematics estimation in natural, uncontrolled environments.

Main Methods:

• A deep learning framework combining neural networks and top-down optimization was developed.
• Deep neural networks were trained on extensive synthetic inertial data from a marker-based motion-tracking database.
• Data augmentation and an automated calibration approach were employed to enhance robustness against sensor placement and limb alignment variability.

Main Results:

• The framework accurately predicted lower extremity joint kinematics, achieving root mean squared errors below 1.27° (flexion/extension), 2.52° (ad/abduction), and 3.34° (internal/external rotation).
• Prediction accuracy improved with increased training data, demonstrating the importance of large datasets for deep learning models.
• The method showed promising results for walking and running trials, with potential for application in uncontrolled settings.

Conclusions:

• The proposed deep learning framework offers a promising, magnetometer-free approach for estimating lower extremity joint kinematics from inertial data.
• This advancement could facilitate large-scale biomechanics studies, providing new insights into disease progression, patient recovery, and sports performance.
• Further validation with real-world IMU data is needed, but the results indicate a significant step towards convenient gait analysis in natural environments.