Validity of multiple human pose estimation tools for measuring knee impact angles in video-captured falls of older adults

Falls are a major cause of injury in older adults. Although bending the knees during a fall has been shown to reduce stress on the hip, knee motion during falls is not well understood because laboratory fall studies are limited by safety concerns and marker occlusion in motion capture systems. AI-based pose estimation may help overcome these challenges, but its accuracy in measuring joint angles during falls has not yet been validated. We evaluated three pose estimation models (OpenPose, VideoPose3D, WHAM) for analyzing knee kinematics in video-captured falls. A total of 121 videos of 13 older adults (64.0 ± 5.9 years) falling sideways, utilizing diverse fall strategies (knee block, stick-like, tuck-and-roll), in a lab setting were analyzed. Each model generated time series of knee angles from the videos, from which knee flexion angles at ground impact were calculated and compared to ground truth data from a motion capture system. Agreement with the ground truth was assessed using mean absolute error (MAE), mean absolute percentage error (MAPE), and bias, analyzed across viewing planes (sagittal vs. frontal) and leg sides (impact vs. opposite). WHAM demonstrated the highest accuracy (MAPE:13.61 ± 10.55%) with minimal bias (