Answers are grounded in this paper and cite their sources.
Two-dimensional (2D) video-based pose estimation offers a low-cost complement to three-dimensional motion capture (MOCAP) for quantifying fall impact kinematics, yet few studies have compared the available tools under controlled conditions. We compared the accuracy and agreement of six 2D pose estimation tools, one manually digitized (Kinovea) and five AI-based (OpenPose, RTMPose, and MediaPipe-lite/full/heavy), against a MOCAP reference. Eighteen healthy young adults fell in three directions under two fall types, with each trial recorded simultaneously by MOCAP and an iPhone. We found that Kinovea showed the highest agreement for peak vertical velocity (ICC = 0.902), while OpenPose and RTMPose showed the lowest error among the AI-based tools. All six tools underestimated peak vertical velocity by 9% on average (13% in MediaPipe-heavy). Errors were smallest in forward falls for both outcomes. Among the AI-based tools, peak vertical velocity errors were greatest in sideways falls for four of five tools, but Kinovea showed its greatest error in backward falls. For horizontal velocity, all six tools showed agreement with MOCAP (ICC = 0.83 to 0.90); no pairwise differences between the AI-based tools were significant after Bonferroni correction. These findings may guide tool selection for fall kinematics research. Further validation is needed for surveillance footage.
No takes yet. Share an insight, caveat, or question.
Lim et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: