Extending 3D markerless tracking for biomechanical analysis of human gait

在我们实验室进行的研究使用数学模型来评估与骨关节炎（OA）相关的运动适应或补偿。一个重要的目标是预测OA的进展，并预测个体OA手术干预的结果。实现个性化护理的这一目标受到以下观察结果的挑战：手术和非手术干预的OA症状、进展和临床结局对个体而言是独特的。具有讽刺意味的是，患者特异性模型要求对大量患者人群进行研究，以便对广泛的属性选择进行分类。 对于我们的研究，三维人体运动的准确测量，以及产生这种运动的潜在力的估计，是使用昂贵的相机来获得的，该相机跟踪放置在身体上的标记的位置，并结合测量脚-地面接触力的传感器。数据收集量小，实验方案耗时且人力密集。对于我们的研究来说，更重要的是，测量的可重复性很差，因为标记放置的可重复性和可靠性很差，这排除了我们实现目标所需的OA大人群研究。 HAS-Motion（我们的工业合作伙伴）开发了一种无标记的3D运动捕捉系统，该系统使用一系列标准摄像机和深度神经网络，在不同的环境中持续快速地记录人体运动。该系统有可能使我们能够进行OA的纵向和多中心研究。在我们采用该系统之前，我们必须在对我们的研究很重要的运动类别的背景下验证该系统的准确性，可重复性和可靠性，并在我们研究中常用的信号和模型的背景下分析和解释数据。我们还建议与HAS-Motion合作，将肌肉骨骼模型和环境约束集成到他们的解决方案中。这个为期三年的项目将支持一个PDF，一个博士生，两个硕士生，一个实验室技术人员和三个本科暑期学生。

The research conducted in our laboratory uses mathematical models for assessing movement adaptations or compensations associated with osteoarthritis (OA). An important goal is to predict the progression of OA, and to predict the outcomes of surgical interventions for OA for an individual. Achieving this goal of personalized care is challenged by the observation that OA symptoms, progression and clinical outcomes from surgical and non surgical interventions are unique to an individual. Ironically patient specific models require that research studies be conducted with large patient populations in order to classify a broad selection of attributes. For our research, accurate measures of three dimensional human motion, and estimation of the underlying forces that generated such motion, are derived using expensive cameras that track the location of markers placed on the body combined with sensors that measure foot-ground contact forces. The data collection volumes are small, and experimental protocols are time consuming and manpower intensive. More importantly for our research the repeatability of measures is poor because the repeatability and reliability of marker placement is poor, which precludes the large population studies of OA that we need to achieve our goals. HAS-Motion (our industrial partner) has developed a markerless 3D motion capture system that uses an array of standard video cameras and deep neural networks to record human motion consistently and quickly across different environments. This system has the potential to allow us to conduct longitudinal and multicentre studies of OA. Before we can adopt this system we must verify the accuracy, repeatability, and reliability of this system in the context of the class of movements that are important to our research, and to analyze and interpret the data in the context of the signals and models that are commonly used in our research. We also propose to collaborate with HAS-Motion to integrate musculoskeletal models and environmental constraints into their solution. This three year project will support one PDF, one PhD student, two Master's students, one lab technician, and three undergraduate summer students.