A Probabilistic Pose Estimation Algorithm for 3D Motion Capture Data

描述（由申请人提供）：康复研究面临的一个主要挑战是测量损伤、功能限制和残疾之间的关系。生物力学分析是建立这些关系的关键工具，通过提供患者状态和治疗结果的定量客观测量。许多生物力学分析的核心是基于使用传感器（光学、电磁或惯性）记录3D运动数据的多段模型的姿势（位置和方向）估计。Visual3D是市面上最先进的用于3D动作捕捉数据的临床生物力学分析软件，它包含了从3D传感器数据中估计姿势的解决方案，这些数据已在世界各地的实验室进行了测试，并在日常临床评估中使用。研究人员开始依赖Visual3D的功能。C-Motion提出了一项合作研究和开发工作，以获得新的姿态估计技术到研究人员手中。第一阶段的算法和第二阶段的增强功能将包含在Visual3D中。Visual3D功能的核心是灵活的算法，用于识别从3D运动捕捉传感器到分割骨骼模型的3D姿态的映射。Visual3D姿态估计算法（以及其他商业生物力学软件）的主要假设是传感器与它们所附着的身体部分刚性移动。然而，可以接受的是，附着在皮肤上的传感器相对于底层骨骼移动，并且软组织人工制品难以量化或建模，因为它通常是系统的，但在个案基础上有所不同。这是对非侵入性临床运动分析相关性的严重挑战。目前的姿态估计算法没有考虑到软组织伪影模型。数据中的不确定性（例如传感器噪声和伪影）不能使用当前的判别方法直接解决，但可以通过在概率推理的一般框架中投射姿态估计问题来解决（Todorov, 2007）。在这个框架中，姿态和任何关于姿态的先验知识都是概率编码的，生成模型捕获“伪影和噪声”，该模型定义给定姿态的数据的条件概率。在第一阶段，我们将实现并测试基于运动学的概率算法，该算法使用Todorov博士提出的贝叶斯推理来计算受试者的姿势（位置和方向）。结果将与我们的合作者Scott Tashman博士（匹兹堡大学生物动力学实验室）同时记录的一组双平面电影透视数据和3D运动捕捉数据进行比较，我们将把这些数据作为骨骼运动的“黄金标准”。整个项目是非常雄心勃勃的，所以在第一阶段，我们正在尝试整个算法的一个重要子集来证明这种方法的可行性，并提供证据，证明我们有能力解决更雄心勃勃的第二阶段项目。

DESCRIPTION (provided by applicant): A major challenge facing rehabilitation research is to measure relationships between impairments, functional limitations, and disabilities. Biomechanical analyses are a key tool for establishing these relationships by providing quantitative objective measures of patient status and treatment outcomes. At the heart of many biomechanical analyses is estimation of the pose (position and orientation) of a multi-segment model based on recording of 3D motion data using sensors (optical, electro-magnetic, or inertial). Visual3D, the most advanced clinical biomechanics analysis software available commercially for 3D motion capture data, contains solutions for the estimation of pose from 3D sensor data that have been tested in laboratories throughout the world, and are used on a daily basis for clinical assessment. Researchers have come to rely on Visual3D's capabilities. C-Motion is proposing a collaborative research and development effort to get new pose estimation techniques into the hands of researchers. The algorithms from Phase I and the enhancements in Phase II will be included in Visual3D. At the core of Visual3D's functionality are flexible algorithms for identifying a mapping from 3D motion capture sensors to the 3D pose of a segmented skeletal model. The principle assumption of the Visual3D pose estimation algorithms (and other commercial biomechanics software) is that sensors move rigidly with the body segments to which they are attached. It is accepted, however, that sensors attached to the skin move relative to the underlying skeleton and that this Soft Tissue Artifact is challenging to quantify or model because it is often systematic but varies on a case by case basis. This artifact is a serious challenge to the relevance of non-invasive clinical motion analyses. The current pose estimation algorithms were not designed to incorporate models of soft tissue artifact. Uncertainty in data (e.g. sensor noise and artifact) cannot be addressed directly using current discriminative methods, but may be addressed by casting the Pose Estimation problem in the general framework of probabilistic inference (Todorov, 2007). In this framework, the pose and any prior knowledge about the pose are encoded probabilistically, and the "artifacts and noise" are captured by a generative model, which defines the conditional probability of the data given the pose. In Phase I we will implement and test a kinematics-based probabilistic algorithm for computing the pose (position and orientation) of a subject using Bayesian inference as proposed by Dr. Todorov. The results will be compared to a set of biplanar cinefluoroscopy data and 3D motion capture data recorded simultaneously by our collaborator Dr. Scott Tashman (Biodynamics Laboratory at the University of Pittsburgh), which we will treat as our "gold standard" for bone motion. The overall project is very ambitious, so in Phase I we are attempting an important subset of the overall algorithm to demonstrate feasibility of this approach, and to provide evidence that we are capable of tackling the even more ambitious Phase II project. PUBLIC HEALTH RELEVANCE: There is a tremendous need for improved rehabilitation research and clinical services to lower individual health care costs and improve productivity and quality of life. Biomechanical analysis is a key tool for understanding the relationships between impairments, functional limitations, and disabilities by providing quantitative, objective measures of patient status and treatment outcomes. This project is designed to apply probabilistic algorithms developed in the field of machine vision to make a new generation of biomechanical techniques available commercially, which will enable researchers to improve movement analysis dramatically and ultimately patient outcomes.