DeepGaitLab: Accurate and Flexible Markerless Motion Tracking Powered by Synthetic Data

Area of Science:

• Biomedical Engineering
• Computer Vision
• Biomechanics

Background:

• Accurate human motion tracking is crucial for diagnosing gait pathologies and optimizing sports performance.
• Current single-view approaches lack accuracy for biomedical applications, while multi-view systems are often complex and require extensive calibration.
• There is a need for scalable, accurate, and user-friendly motion tracking solutions for both research and clinical settings.

Purpose of the Study:

• To introduce DeepGaitLab, an open-source framework for accurate 3D human motion tracking.
• To overcome the limitations of existing motion tracking systems, particularly in terms of accuracy, calibration requirements, and generalizability.
• To provide a practical and scalable platform for real-world deployment in biomechanics and clinical applications.

Main Methods:

• Developed DeepGaitLab, an open-source framework utilizing synthetic data for training.
• Enabled flexible camera configurations, including setups with as few as two cameras, eliminating the need for inter-camera calibration.
• Implemented an optional environment-specific fine-tuning strategy to enhance accuracy without requiring new labeled data.

Main Results:

• DeepGaitLab achieves accurate 3D kinematics, outperforming commercial and open-source alternatives.
• Accuracy shows monotonic improvement with the addition of more cameras.
• Performance remains robust for individuals with mobility limitations, including those recovering from anterior cruciate ligament reconstruction, stroke, and with mild cognitive impairment.
• Demonstrated utility across three distinct clinical populations in two different environments.

Conclusions:

• DeepGaitLab offers a practical and scalable solution for accurate 3D human motion tracking.
• The framework bridges the gap between research-grade biomechanics, AI tools, and clinical applications.
• Publicly sharing data, code, and models promotes accessibility and further innovation in human movement analysis.