3D pose estimation for scalable remote gait kinematics assessment.

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Computer Vision

Background:

• Marker-based gait analysis is limited by cost and clinical settings.
• Continuous, real-world monitoring is crucial for neurological rehabilitation, especially for spinal cord injury (SCI).
• Existing methods lack scalability and accessibility for remote patient monitoring.

Purpose of the Study:

• To develop and validate a scalable, markerless gait analysis system for pathological gait in SCI.
• To benchmark state-of-the-art 3D pose estimation models for accuracy in pathological gait analysis.
• To establish a non-invasive framework for remote, longitudinal monitoring of gait in SCI patients.

Main Methods:

• Benchmarking of 3D pose estimation models, identifying VideoPose3D as optimal.
• Development of an automated pipeline using a single video for kinematic data extraction.
• Analysis of time-series joint angle data from the largest SCI gait dataset (225 participants).

Main Results:

• VideoPose3D demonstrated superior performance for pathological gait analysis in SCI.
• The automated pipeline successfully extracted detailed kinematic data and identified distinct gait patterns.
• Key gait biomarkers, including reduced hip/knee flexion, were identified in the SCI cohort.

Conclusions:

• Markerless pose estimation is clinically viable for pathological gait analysis.
• The developed framework enables non-invasive, scalable, and remote monitoring of gait.
• This approach facilitates personalized rehabilitation and supports a preventative medicine strategy for SCI.