Accuracy of Computer Vision-Based Pose Estimation Algorithms in Predicting Joint Kinematics During Gait

Abstract Accurate and user-friendly joint kinematic measurement is essential in clinical gait assessment. Pose estimation algorithms offer an alternative to cumbersome marker-based motion capture, whether optical or IMU-based. This study assesses AlphaPose and BlazePose pose estimation tools efficacy in determining gait joint kinematics against Vicon motion capture. Ten healthy male participants walked at varied speeds, with concurrent measurements via thirty-nine reflective skin markers and a GoPro RGB camera in sagittal plane. Pose estimation tools processed videos, and trigonometric calculations derived joint angles. BlazePose demonstrated lower root mean square error (RMSE) values than Vicon, with a maximum of 14.2° in the left knee during slow gait. The Wilcoxon signed-rank test revealed significant joint measurement differences, worsening with speed. Both AlphaPose and BlazePose differ in performance compared to Vicon. AlphaPose generally yielded higher ranges of motion (RoM) and larger RMSE values, while BlazePose exhibited elevated normalized RMSE values. With further improvements to BlazePose algorithm, clinicians will be empowered to conduct real-time pre- and post-intervention gait assessments effortlessly using standard cameras.