Study of Kinematic Tracking and Monitoring of Human Movements in a Collaborative Network of Depth Sensors

通过观察运动模式来跟踪人体运动是一个具有许多潜在应用的基础研究领域，例如人类步态分析，独立生活设施中的老年人监控，人机交互和监控。迄今为止，相机或可穿戴传感器的网络已被用于捕获人体运动。然而，视觉感测一直在努力克服照明和表面纹理特性的变化，并且可穿戴传感器面临较差的接受度。目前正在开发一种更可行和更经济的方法，使用低成本的动态深度传感器，这是一种新兴技术。拟议的研究计划将针对通过分布式深度传感器网络对人进行鲁棒运动跟踪的方面。研究成果将有助于人类生物力学和机器人领域，并以多种方式使加拿大人受益。该研究将提供一种实用的工具，可用于患者康复，通过观察他们在自然生活环境中的运动而无需佩戴传感器的不便;它还将提供一种独特的非侵入性方法，用于监测我们老龄化人口的活动，以确保他们在私人或公共设施中的安全。这项研究将为各种深度传感技术网络开发新的校准方法，建模和理解与身体位置和肢体运动相关的测量的噪声和灵敏度的性质。提出了基于深度测量的两种运动学跟踪模型：粗跟踪模型和细跟踪模型。在粗模型中，我们定义的整体周围的人的形状的基础上的点在四肢和新的方法的基础上的形状的横截面切割。我们建议扩展将物理监测区域划分为较粗体积的概念（例如，立方体）并将分布式深度测量与对应的3D体积相关联。对于每一个体积，我们将探索各种方法来找到一个合适的表示所规定的测量深度信息的表面。然后，重建的粗略形状模型被用作跟踪人的所选肢体（例如，手臂、脚和头部）的基础。首先，关于末端的位置的信息被用来定义沿它们之间的网格模型的局部距离函数沿着。对于占据一组立方体的每个肢体，将定义跟踪变量以表示肢体的底层骨架和局部简单几何形状。由于自然的不确定性与身体和肢体运动，每个肢体的跟踪方法提出了一种新的智能滤波器框架（智能粒子滤波器）的基础上。我计划开发，研究和实验跟踪变量的各种运动模型和先验运动概率分布，这些模型可以代表每个时间步的预期跟踪变量的知识。然后，将定义一组预测的跟踪变量，其可以与实际测量的深度传感器信息进行比较和加权。预期的新的贡献是与一个强大的基于模型的切换方法的开发跟踪作为一个功能的人的全局运动意图和所选肢体的局部运动模式。总体目标是从跟踪一个人及其相关肢体开始，并将结果扩展到在监测区域内移动的多个人。在开发的每个阶段，增量结果将针对现有的基于标记的系统进行验证，并与其他已知的运动预测方法进行比较。

Tracking human body movement by observing motion patterns is a fundamental research area with many potential applications, such as human gait analysis, monitoring seniors in independent living facilities, human-robot interaction, and surveillance. To date, networks of cameras or wearable sensors have been utilized to capture human body movements. However, visual sensing has struggled to overcome variations in illumination and surface texture properties, and wearable sensors have faced poor acceptance. A more feasible and economical approach is currently being developed using low cost kinematic depth sensors, which are an emerging technology. The proposed research program will target the aspects of robust motion tracking of people through a network of distributed depth sensors. Research results will contribute to the field of human biomechanics and robotics and benefit Canadians in a number of ways. The research will provide a practical tool that can be used in patient rehabilitation by observing their movements in their natural living habitat without the inconvenience of wearing sensors; it will also offer a unique, non-intrusive approach for their deployment in monitoring activities of our aging population for their safety in private or public caregiving facilities. This investigation will undertake development of novel calibration methods for various networks of depth sensing technologies, modeling and understanding the nature of noise and sensitivity of measurements related to the location of bodies and movements of limbs. Two kinematic tracking models are proposed based on the depth measurements: a coarse tracking model; and a fine tracking model. In the coarse model, we define the overall surrounding shape of persons based on points at extremities and the novel method based on shapes of cross-sectional cuts. We propose to extend the notion of dividing the physical monitoring area into coarser volumes (e.g., cubes) and associate the distributed depth measurements to corresponding 3D volumes. For each volume, we will explore various approaches for finding a suitable representation of the surface prescribed by measured depth information. The reconstructed coarse shape model is then used as a basis for tracking selected limbs of the person, e.g. arms, feet, and head. First, information about the location of extremities is used to define a local distance function along the mesh model between them. For each limb occupying a set of cubes, tracking variables will be defined to represent the underlying skeleton and local simple geometrical shape of the limb. Due to natural uncertainties associated with body and limb movements, a tracking method for each limb is proposed based on a novel intelligent filter framework (intelligent particle filter). I plan to develop, study, and experiment with various motion models of the tracking variables and prior motion probability distributions that can represent the knowledge of expected tracking variables at each time step. Then a set of predicted tracking variables will be defined that can be compared and weighted with the actual measured depth sensor information. The expected novel contributions are associated with development of a robust model-based switching method for tracking as a function of global motion intentions of the person and the local motion patterns of the selected limbs. The overall objective is to start by tracking one person and their associated limbs and extend the results to multiple people moving in the monitoring area. At each stage of the development, incremental results will be validated against an existing marker-based system and compared with other known motion prediction methods.