Evaluation of CNN-Based Human Pose Estimation for Body Segment Lengths Assessment

Abstract

Human pose estimation (HPE) methods based on convolutional neural networks (CNN) have demonstrated significant progress and achieved state-of-the-art results on human pose datasets. In this study, we aimed to assess the performance of CNN-based HPE methods for measuring anthropometric data. A Vicon motion analysis system as the reference system and a stereo vision system recorded ten asymptomatic subjects standing in front of the stereo vision system in a static posture. Eight HPE methods estimated the 2D poses which were transformed to the 3D poses by using the stereo vision system. Percentage of correct keypoints, 3D error, and absolute error of the body segment lengths are the evaluation measures which were used to assess the results. Percentage of correct keypoints – the standard metric for 2D pose estimation – showed that the HPE methods could estimate the 2D body joints with a minimum accuracy of 99%. Meanwhile, the average 3D error and absolute error for the body segment lengths are 5 cm.