Using computer vision and deep learning to measure worker kinematics

项目摘要/摘要 肌肉骨骼疾病(MSD)是最常见和最昂贵的非致命性工伤之一， 疾病几乎遍及美国所有的工业部门。回应NIOSH中强调的明确需求 国家肌肉骨骼健康研究议程，以制定改进的接触估计方法 MSD的职业风险因素，这项研究将验证新的测量工人姿势和 仅使用标准视频作为输入的动作。该软件充分利用了计算机视觉领域的重大进步 和机器学习科学，直到最近才能在三个月内测量人体姿势 使用标准二维视频或图像源的多维空间。最终，我们的长期计划之一 目标是为职业安全和健康从业者开发类似于广泛使用的直接应用程序 评估职业危害暴露的读数仪器(如声压计和气体 监视器)。在最初的R21中，我们建议验证我们的软件产生的姿势数据(目标1)和 检查我们软件输出的姿势信息与更传统的软件输出的姿势信息之间的一致性 (但耗时的)基于观察的视频分析(目标2)。在目标1中，参与者将执行重复的、 手臂密集型任务，包括伸手和操纵安装在固定装置前面的按钮 尸体。然后我们将评估颈部、肩部、肘部、腕部、躯干和膝部角度的准确性。 通过我们的软件测量的位移(即随时间的姿势)，与同时收集的数据进行比较 使用光学运动捕捉系统。实验变量包括所需的运动范围 参与者执行任务以及用于录制参与者视频的摄像头的配置 这项任务。AIM 1的结果将提供有关我们所需的新软件性能的关键信息 为能够实地实施的暴露评估应用程序提供最佳做法。在目标2中，我们 将重新分析在上一次前瞻性研究过程中获得的&gt；1000个工作场所视频 制造业工人中的极端MSD。提出了评估方法间协议的分析方法 在自动化视频分析(我们的软件)和由训练有素的专家观察员完成的分析之间 在前瞻性研究的过程中。结果将提供证据，证明我们的软件可以量化 职业暴露于MSD危险因素的时间仅为进行常用观察所需时间的一小部分- 基于分析。对现有工作场所视频的重新分析也可以开辟新的探索途径 职业暴露于MSD危险因素与未来事件健康结局的关系 学习。

PROJECT SUMMARY/ABSTRACT Musculoskeletal disorders (MSDs) are among the most frequent and costly nonfatal work-related injuries and illnesses across virtually all US industry sectors. Responding to the clear need emphasized in the NIOSH National Research Agenda for Musculoskeletal Health to develop improved methods of estimating exposure to occupational risk factors for MSDs, this research will validate new software for measuring worker postures and movements using only standard video as input. The software leverages major advances in computer vision and machine learning sciences that only recently have enabled measurement of human postures in three dimensional space using standard two dimensional video or image sources. Ultimately, one of our long-term goals is to develop applications for occupational safety and health practitioners analogous to widely-used direct reading instruments for assessing exposure to occupational hazards (e.g., sound pressure meters and gas monitors). In this initial R21, we propose to validate the postural data our software produces (Aim 1) and examine agreement between postural information output by our software and that output by more traditional (but time-consuming) observation-based video analyses (Aim 2). In Aim 1, participants will perform a repetitive, arm-intensive task involving reaching to and manipulating knobs mounted to a fixture located in front of the body. We will then estimate the accuracy of neck, shoulder, elbow, wrist, trunk, and knee angular displacements (i.e., posture over time) measured by our software, compared to data simultaneously collected using an optical motion capture system. Experimental variables include the range of motion required of participants to perform the task and the configuration of the camera used to record video of participants during the task. Results from Aim 1 will provide critical information about the performance of our new software needed to inform best-practices for implementation in field-capable exposure assessment applications. In Aim 2, we will reanalyze >1000 workplace videos obtained during the course of a previous prospective study of upper extremity MSDs among manufacturing workers. Analyses are proposed to assess the inter-method agreement between automated video analyses (our software) and analyses completed by trained specialist observers during the course of the prospective study. Results will provide evidence that our software can quantify occupational exposure to MSD risk factors at a fraction of time needed to perform commonly used observation- based analyses. The reanalysis of existing workplace videos can also open new pathways to explore associations between occupational exposures to MSD risk factors and incident health outcomes in future studies.