Verification of reliability and validity of motion analysis systems during bilateral squat using human pose tracking algorithm

Abstract

Background: The human tracking algorithm called OpenPose can detect joint points and calculate joint angles. However, the reliability and validity of OpenPose have not been clarified yet. Research question: Are there the enough reliability and validity of OpenPose based motion analysis? Methods: 20 healthy young subjects participated in this study. The motion task was a bilateral squat. The joint angles of the trunk, hip, knee, and ankle were calculated using OpenPose and VICON. Kinematic measurements by three-dimensional motion analysis devices were recorded using VICON. Simultaneously, the images were taken with a digital camera from the right side. After the images were processed with OpenPose, joint angles were calculated from estimated joint points. To confirm the test-retest reliability within device, intraclass correlation coefficients [ICC (1, 3)] were calculated. To confirm the validity, linear regression analysis and ICC (2, 1) between the data obtained by OpenPose and VICON were calculated. Furthermore, the agreement between the data obtained by OpenPose and VICON was assessed by Bland-Altman analysis. Results: ICCs (1, 3) of the data obtained by OpenPose and VICON were almost perfect. There were significant associations between the data obtained by OpenPose and VICON. ICCs (2, 1) between the data obtained by OpenPose and VICON were almost perfect or substantial for trunk, knee and ankle joints, and fair on the hip joint. There were fixed biases on knee and ankle joints, and proportional biases on trunk and hip joint. Significance: OpenPose based motion analysis is reliable and has the advantage of being low cost and easier to operate than conventional methods. In future, to consider the clinical utility of OpenPose, it is necessary to identify the error between the true values indicating actual joint movement and data obtained by OpenPose with its correction for fixed and proportional biases.