3D pose estimation for scalable remote gait kinematics assessment.

Abstract

Marker-based gait analysis is a cornerstone of neurological rehabilitation, but its high cost and clinical setting dependency limit its use for continuous, real-world monitoring. This study presents a scalable, markerless approach for pathological gait analysis, addressing a critical need for remote monitoring in conditions such as spinal cord injury (SCI). To ensure the highest fidelity for this application, we first investigated and benchmarked state-of-the-art 3D pose estimation models, finding VideoPose3D to be the best-performing for pathological gait in SCI. Our automated pipeline utilizes this model with a single video to generate detailed kinematic data, extracting time-series data for key joint angles. Our approach automatically identifies distinct gait patterns in the largest SCI dataset collected to date, which includes 225 participants. Specifically, key biomarkers such as reduced hip/knee flexion during phases of the gait cycle. This work establishes the clinical viability of markerless pose estimation for pathological gait analysis, offering a non-invasive and scalable framework to enhance personalized rehabilitation and enable longitudinal condition monitoring outside the clinic. Our approach provides clinically relevant metrics needed to translate real-world data into actionable treatments, paving the way for a preventative medicine approach.