Learning-based 3D human kinematics estimation using behavioral constraints from activity classification.

Area of Science:

• Biomechanics
• Machine Learning
• Wearable Technology

Background:

• Inertial measurement units (IMUs) provide a portable, cost-effective alternative to lab-based motion capture.
• IMU-based motion tracking faces challenges from signal drift errors amplified by numerical integration.
• Existing drift reduction methods often require body parameter measurements or lack accuracy for diverse applications.

Purpose of the Study:

• To develop an end-to-end machine learning model for enhanced kinematics estimation using two IMUs.
• To incorporate human behavioral constraints and activity classification into the estimation process.
• To improve the accuracy of joint angle and movement trajectory measurements.

Main Methods:

• Introduced the Activity-in-the-loop Kinematics Estimator, an integrated machine learning model.
• Leveraged human behavioral constraints and activity classification for kinematics estimation.
• Utilized two IMUs for motion tracking in dynamic scenarios.

Main Results:

• Achieved trajectory errors under 0.021 m and shoulder joint angle errors below $$.
• Demonstrated a 52% reduction in trajectory error and a 17% reduction in shoulder joint angle error compared to models without activity classification.
• Validated accurate motion tracking with minimal IMUs and domain-specific context.

Conclusions:

• The Activity-in-the-loop Kinematics Estimator significantly enhances kinematics estimation accuracy.
• Integrating activity classification with behavioral constraints improves IMU-based motion tracking performance.
• This approach offers accurate motion tracking with minimal hardware and contextual information.