CAREER: Bayesian Graph Signal Processing for Machine Perception

机器感知的算法方法可以检测、定位和跟踪环境中的物体。他们将在海洋科学、机器人、自动驾驶、室内定位和人群计数等领域建立新的服务和应用。现有的方法依赖于简化和预处理阶段，降低了测量的数据速率，但也丢弃了相关信息，从而限制了性能。特别是，如果物体彼此靠近，或者它们产生的测量值很弱，现有的方法通常无法可靠地感知它们。该项目旨在建立新的感知方法，优化利用所有可用信息，在具有挑战性的场景中提供前所未有的性能。能够使用大量具有高数据速率的传感器的关键原理是系统地利用感知问题的数学公式中的图结构。开发的方法将使用矢量传感器和大型水听器阵列提供的水声数据进行评估。该项目带来的创新将大大提高海洋感知系统的性能，同时也将在自动驾驶、医疗成像和无线通信等各种进一步应用领域取得切实进展。跨学科教育和推广活动旨在使不同的学生群体接触到最先进的机器学习和感知技术，以及它们在海上的应用。研究成果将分发给科学界，并用于教材以及将在全世界提供的教程和短期课程。这个项目将引入基于图的估计来建立感知方法，利用所有可用的信息，从而产生前所未有的感知性能。利用“拉伸”或“打开”图节点的原理，将高维操作替换为低维操作。基于这一原理，将开发出具有收敛性保证、大大降低计算复杂度和优越可扩展性的迭代感知方法。与传统的对象感知方法相反，所设想的方法的高可扩展性使得生成和维护非常大量的对象假设成为可能，从而提高感知性能。特别有趣的是为每个实值或复值数据单元（样本、像素或bin）形成新的对象假设的方法，并且每个数据单元在基于整体图的框架中与对象假设概率相关。此外，研究团队将引入基于图的估计方法，该方法嵌入物理环境模拟器，利用多路径传播和虚拟孔径，以提高对低可观察对象的感知，并提供对不确定环境参数的鲁棒性。该项目还将设计基于图的机器感知方法的扩展，通过从数据中学习的信息自适应地改进底层统计模型。在这里，描述感知问题的原始统计模型的图由一个图形神经网络支持，该网络使用标记的真实数据或物理环境模拟器提供的合成数据进行训练。最后，一个用于展示感知能力的开放软件和硬件平台将把研究和教育结合起来，并支持全面的推广和传播活动。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Algorithmic methods for machine perception can detect, localize, and track objects in the environment. They will establish new services and applications in fields such as ocean sciences, robotics, autonomous driving, indoor localization, and crowd counting. Existing methods rely on simplifications and preprocessing stages that reduce the data rate of the measurements, but also discard relevant information and thus limit performance. In particular, if objects are close to each other or the measurements they generate are weak, existing methods are often unable to perceive them reliably. This project aims to establish new perception methods that make optimal use of all the available information to provide unprecedented performance in challenging scenarios. The key principle that will enable the use of a large number of sensors with high data rates, is to systematically exploit graph structures in the mathematical formulation of perception problems. The developed methods will be evaluated using underwater acoustic data provided by vector sensors and large arrays of hydrophones. Innovation resulting from this project will substantially improve the performance of marine perception systems but also lead to tangible advances in a variety of further applications including autonomous driving, medical imaging, and wireless communication. Interdisciplinary education and outreach activities aim to expose a diverse cohort of students to state-of-the-art machine learning and perception techniques as well as their deployment at sea. Research results will be disseminated to the scientific community and used in teaching materials as well as tutorials and short courses to be presented worldwide.This project will introduce graph-based estimation to establish perception methods that make use of all the available information and thus yield unprecedented perception performance. The principle of "stretching" or "opening" graph nodes will be employed to replace high-dimensional operations by lower-dimensional ones. Based on this principle, iterative perception methods with convergence guarantees as well as strongly reduced computational complexity and superior scalability will be developed. Contrary to conventional object perception approaches, the high scalability of the envisaged methods makes it possible to generate and maintain a very large number of object hypotheses and, in turn, improve perception performance. Of particular interest are methods where a new object hypothesis is formed for each real- or complex-valued data cell (sample, pixel, or bin), and each data cell is probabilistically associated with an object hypothesis in a holistic graph-based framework. In addition, the research team will introduce graph-based estimation methods that embed simulators of the physical environment to exploit multipath propagation and virtual apertures with the goal of improving the perception of low-observable objects and providing robustness against uncertain environmental parameters. The project will also devise an extension of graph-based machine perception methods that adaptively refines the underlying statistical model by information learned from data. Here, the graph that describes the original statistical model of the perception problem is supported by a graph neural network trained with labeled real data or with synthesized data provided by simulators of the physical environment. Finally, an open software and hardware platform for the demonstration of perception capabilities will bring together research and education as well as support comprehensive outreach and dissemination activities.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.