Michaels Reese, Barreira Tiago V, Robinovitch Stephen N, Sosnoff Jacob J, Moon Yaejin
Department of Exercise Science, Syracuse University, 150 Crouse Dr, Syracuse, NY, 13244, USA.
Department of Biomedical Physiology and Kinesiology, Simon Fraser University, Burnaby, BC, Canada.
Sci Rep. 2025 Jan 10;15(1):1558. doi: 10.1038/s41598-025-85934-y.
Analyzing video footage of falls in older adults has emerged as an alternative to traditional lab studies. However, this approach is limited by the labor-intensive process of manually labeling body parts. To address this limitation, we aimed to validate the use of the AI-based pose estimation algorithm (OpenPose) in assessing the hip impact velocity and acceleration of video-captured falls. We analyzed 110 videos of 13 older adults (64.0 ± 5.9 years old) falling sideways in an experimental setting. By applying OpenPose to each video, we generated a time series of hip positions in the video, which were then analyzed using custom MATLAB code to calculate hip impact velocity and acceleration. These calculations were compared against ground truth measurements obtained from motion capture systems (VICON for hip impact velocity) and inertial measurement units (MC10 for hip impact acceleration). We examined the agreement between the ground truth and OpenPose measurements in terms of mean of absolute error (MAE), mean of absolute percentage error (MAPE), and bias (mean of error). Results showed that OpenPose had a good accuracy in estimating hip impact velocity with minimal bias (MAE: 0.17 ± 0.13 m/s, MAPE: 7.28 ± 5.21%; percent bias: - 1.27%). However, its estimation of hip impact acceleration (i.e., peak vertical hip acceleration at impact) showed poor accuracy (MAPE: 26.3 ± 19.4%), showing substantial underestimation in instances of high acceleration impacts (> 3.0 g). Further ANOVA analysis revealed OpenPose's ability to discern significant differences in hip impact velocity and acceleration based on the movement response utilized during the fall (e.g., stick-like fall, tuck-and-roll, knee block). This is the first study to validate the use of a pose estimation algorithm for identifying the hip impact kinematics in video-captured falls among older adults. Future validation studies involving diverse camera settings, fall contexts, and biomechanical parameters are warranted to extend this support for using pose estimation algorithms in this field.
分析老年人跌倒的视频片段已成为传统实验室研究的一种替代方法。然而，这种方法受到手动标记身体部位的劳动密集型过程的限制。为了解决这一限制，我们旨在验证基于人工智能的姿态估计算法（OpenPose）在评估视频捕捉到的跌倒中髋部撞击速度和加速度方面的应用。我们分析了13名老年人（64.0±5.9岁）在实验环境中侧向跌倒的110个视频。通过将OpenPose应用于每个视频，我们生成了视频中髋部位置的时间序列，然后使用自定义的MATLAB代码对其进行分析，以计算髋部撞击速度和加速度。将这些计算结果与从运动捕捉系统（用于髋部撞击速度的VICON）和惯性测量单元（用于髋部撞击加速度的MC10）获得的地面真值测量结果进行比较。我们从绝对误差均值（MAE）、绝对百分比误差均值（MAPE）和偏差（误差均值）方面检查了地面真值与OpenPose测量结果之间的一致性。结果表明，OpenPose在估计髋部撞击速度方面具有良好的准确性，偏差最小（MAE：0.17±0.13 m/s，MAPE：7.28±5.21%；百分比偏差：-1.27%）。然而，其对髋部撞击加速度（即撞击时髋部垂直峰值加速度）的估计准确性较差（MAPE：26.3±19.4%），在高加速度撞击（>3.0 g）的情况下存在明显低估。进一步的方差分析揭示了OpenPose能够根据跌倒过程中使用的运动反应（例如，直挺式跌倒、翻滚、膝盖阻挡）辨别髋部撞击速度和加速度的显著差异。这是第一项验证使用姿态估计算法识别老年人视频捕捉到的跌倒中髋部撞击运动学的研究。未来有必要进行涉及不同相机设置、跌倒情境和生物力学参数的验证研究，以扩展对在该领域使用姿态估计算法的支持。
