Reliability and accuracy of 2D lower limb joint angles during a standing-up motion for markerless motion analysis software using deep learning

Abstract

OpenPose (OP) and DeepLabCut (DLC) are applications that use deep learning to estimate posture, but there are few reports on the reliability, validity, and accuracy of their 2D lower limb joint motion analysis. This study compared OP and DLC estimates of lower extremity joint angles in standing movements with those of conventional software. A total of nine healthy men participated. The trial task was to stand up from a chair. The motion was recorded by a digital camera, and the joint angles of the hip and knee joints were calculated from the video using OP, DLC, and Kinovea. To confirm reliability and validity, ICC was calculated using the Kinovea value as the validity criterion and the correlation coefficient between OP and DLC. In addition, the agreement between those data was evaluated by the Bland-Altman plot. To evaluate the accuracy of the data, root means square error (RMSE) was calculated and compared for each joint. Although the correlation coefficients and ICC (2, 1) were in almost perfect agreement, fixed and proportional errors were found for most joints. The RMSE was smaller for OP than for DLC. Compared to Kinovea, OP and DLC can estimate the joint angles of the hip and knee joints during the stand-up movement with an estimation error of fewer than 10°, but since they are affected by the resolution of the analysis video and other factors, they need to be validated in a variety of environments and with a variety of movements.