Antipodal Video-Based Estimation of Absolute 3D Motion for In-Home Spatio-Temporal Gait Parameter Assessment

Spatio-temporal gait parameters are crucial indicators for analyzing gait, where vision-based methods are widely used recently owing to their low cost and ease of deployment. However, such methods still face challenges in accurately recovering absolute 3D dynamics and robust gait cycles from monocular 2D or 2.5D estimations, and generally require professional supervision to deploy in practice. Here, we propose a framework that estimates all components required for calculating spatio-temporal gait parameters directly from dual-view RGB video sequences under an instructed antipodal deployment setup, enabling self-deployable gait parameter assessment for patients. We design a dual-view 3D keypoint estimation network that decouples absolute 3D keypoint estimation into two steps: first estimating relative 3D keypoints, then regressing the absolute trajectory of a reference keypoint. This separation allows each subtask to be effectively addressed with the available information, whereas information may be insufficient for solving them jointly. Gait cycle segmentation is performed on the estimated relative keypoints with a semi-supervised model to reduce labeling cost and improve robustness to low-quality input. Extensive experiments are conducted to evaluate the performance of each module and the framework, which are validated on a shared dataset split comprising variations in camera placements, subjects, and recording sessions. The whole framework achieves an average relative error of 8.54% across all gait parameters, evaluated over 22 standard parameters commonly provided by motion capture, with only 3.91% for temporal parameters, specifically. This work provides a practical solution for patient deployable, in-home applicable, and accurate gait parameter assessment for long-term tracking and monitoring.