Comparing the Quality of Human Pose Estimation with BlazePose or OpenPose

Abstract

Human pose estimation is a computer vision task that predicts the position of person's body landmarks within a given image or video. This technology could help provide virtual motion assessments by analyzing videos captured when the patient is outside a clinical setting. In this study, a newer pose estimation model that can run on a smartphone (BlazePose) was compared to a well-accepted solution (OpenPose) to determine if these models can provide clinically viable body keypoints for virtual motion assessment. Using ten videos of clinically relevant movements (recorded by physicians), keypoint coordinates were generated from each model. Using OpenPose as a baseline, Pearson correlation and root mean square error were calculated between the BlazePose and OpenPose keypoint trajectories. BlazePose had more instances where keypoints deviated from anatomical joint centres, compared to OpenPose, indicating the BlazePose was not yet viable for clinically relevant assessments. However, BlazePose runtime was much faster than OpenPose and returned metrics that could be incorporated into a smartphone solution. Future designs of a smartphone-based system for conducting virtual motion assessments should utilize OpenPose for pose estimation; however, BlazePose could be used for other design aspects such as movement pre-screening or activity classification.