High-dimensional estimation for sensor systems: from breast cancer detection to autonomous cars

传感器系统的高维估计：从乳腺癌检测到自动驾驶汽车

• 批准号: RGPIN-2017-04269
• 负责人: Coates, Mark
• 金额: $ 2.7万
• 依托单位: McGill University
• 依托单位国家: 加拿大
• 项目类别: Discovery Grants Program - Individual
• 财政年份: 2019
• 资助国家: 加拿大
• 起止时间: 2019-01-01 至 2020-12-31
• 项目状态: 已结题
• 来源: https://www.nserc-crsng.gc.ca/ase-oro/Details-Detailles_eng.asp?id=686941
• 关键词: dimensional; estimation; sensor; systems; breast

Sensor systems are becoming ubiquitous as we strive to improve our understanding of our surroundings and ourselves. Extensive sensor networks are being deployed to monitor the environment. People have started to wear bio-sensing devices to track their health status and physical activity. Self-driving cars are becoming a reality. In many sensor systems, we are faced with a deluge of measurements and must estimate a high-dimensional state that summarizes the pertinent information. For example, a self-driving car needs to locate and track all other vehicles and pedestrians, and must identify and recognize road conditions such as traffic lights. ***Although there have been important advances, we lack algorithms and computing infrastructure that can efficiently process a high-volume stream of measurements and perform high-dimensional state estimation. Modern sensor networks can involve hundreds of sensor nodes, each generating hundreds of measurements per second. To describe the underlying state we may need hundreds of dimensions. Consider a "traffic state vector" describing the evolution of hundreds of traffic flows in a city, where the measurements are the counts of cars traversing all major intersections every second. The challenge intensifies when we add the requirement of real-time processing, which is critical for acquiring an evolving understanding of the status of the environment or system so that we can react accordingly. Many of the current algorithms cannot achieve sufficient accuracy; others require too much computation and cannot operate in real-time. ***I will develop novel algorithms that can estimate and predict the state of a system, in real time, when the state dimension is in the hundreds and we must process thousands of measurements per second. The algorithms will be based on sequential Monte Carlo methods and particle flow. Concepts from island particle filters will be used to parallelize the algorithms, making it feasible to employ them in real-time tracking applications. I will extend these algorithms to the case of multi-sensor, multi-object tracking, where we must also determine how many objects are present in a scene. ***I will apply the algorithms in two application domains: breast cancer detection using radio-frequency (RF) measurements and environmental perception for self-driving vehicles. We have developed a prototype radio-frequency breast cancer detection system and the research in this program will provide the detection algorithms needed to process the data obtained from a scan to decide if a tumour is present. For self-driving vehicles, the multi-sensor, multi-object tracking algorithms will allow us to process the measurements from multiple high-resolution radar sensors to determine in real-time the presence, locations, and shapes of other vehicles. These tracking algorithms will be provided to collaborators and integrated into autonomous driving systems.

随着我们努力提高对周围环境和自身的理解，传感器系统变得无处不在。正在部署广泛的传感器网络来监测环境。人们已经开始佩戴生物传感设备来跟踪他们的健康状况和身体活动。自动驾驶汽车正在成为现实。在许多传感器系统中，我们面临着大量的测量，必须估计一个高维的状态，总结相关的信息。例如，自动驾驶汽车需要定位和跟踪所有其他车辆和行人，并且必须识别和识别交通信号灯等道路状况。* 虽然已经取得了重要的进展，但我们缺乏算法和计算基础设施，可以有效地处理大量的测量数据流并执行高维状态估计。现代传感器网络可能涉及数百个传感器节点，每个节点每秒产生数百个测量值。为了描述潜在的状态，我们可能需要数百个维度。考虑一个描述城市中数百个交通流演变的“交通状态向量”，其中测量值是每秒穿过所有主要十字路口的汽车的计数。当我们增加实时处理的要求时，挑战就会加剧，这对于获得对环境或系统状态的不断了解至关重要，以便我们能够做出相应的反应。目前的许多算法不能达到足够的精度;其他算法需要太多的计算，不能实时操作。* 我将开发新的算法，可以估计和预测系统的状态，在真实的时间，当状态维度是在数百个，我们必须处理数千个测量每秒。该算法将基于顺序蒙特卡罗方法和粒子流。岛粒子滤波器的概念将被用来并行化的算法，使其可行的实时跟踪应用程序中使用它们。我将把这些算法扩展到多传感器、多对象跟踪的情况，在这种情况下，我们还必须确定场景中存在多少个对象。*** 我将在两个应用领域应用这些算法：使用射频（RF）测量的乳腺癌检测和自动驾驶车辆的环境感知。我们已经开发了一个原型射频乳腺癌检测系统，该计划的研究将提供处理扫描获得的数据所需的检测算法，以确定是否存在肿瘤。对于自动驾驶车辆，多传感器、多目标跟踪算法将允许我们处理来自多个高分辨率雷达传感器的测量结果，以实时确定其他车辆的存在、位置和形状。这些跟踪算法将提供给合作者，并集成到自动驾驶系统中。