Collaborative Research: CNS Core: Medium: Network-Enabled Cooperative Perception for Future Autonomous Vehicles

每年，全球道路交通事故造成约135万人死亡。自动驾驶和驾驶辅助技术可以大大提高车辆的安全性。然而，对于这些技术来说，确保在广泛的异常流量情况下的可靠性仍然是一个关键的挑战。为了被社会接受，这些技术必须达到或超过人类驾驶安全水平（例如死亡人数之间的1亿英里）。如今，每辆自动驾驶汽车都依靠自己的传感器来感知环境，并做出独立的驾驶决策。不幸的是，这些传感器依赖于视线感知，车辆的视野可能会被其他车辆阻挡。借助5G和高通（Qualcomm）的蜂窝v2x技术等先进的无线通信技术，车辆将能够直接或利用路边基础设施与其他车辆共享传感器数据，从而有效地看穿障碍物，我们将这种能力称为网络支持的协同感知。虽然网络支持的协同感知是一项引人注目的技术，但网络仍然是一个基本瓶颈。目前的车载通信技术在实践中可以达到6- 10mbps，但先进的车载传感器可以产生数百Mbps的原始数据。该项目旨在解决丰富的原始传感器数据与网络瓶颈之间的紧张关系，同时将网络支持的协作感知扩展到具有多模式交通的极其密集的交通情况，如行人、自行车、三轮车、卡车、汽车等，其中并非所有参与者都可能配备传感器。该项目将开发用于大规模网络协同感知的抽象、算法和工具，以一种称为一瞥的抽象为基础，这是对车辆传感器视图的一部分进行处理的表示。瞥见可以表示视图中的单个对象，或3D空间中的网格，并且可以以不同的粒度表示，以权衡带宽的细节。考虑到这种抽象，该项目将开发：在复杂和高度动态的交通环境中，确定哪些车辆和何时需要哪些一瞥表示的方法；在尊重信道容量限制的情况下，协调这些信息向车辆传输的调度算法；使用增强复合视图训练机器学习模型以做出控制决策的方法；以及确保对一瞥中毒的稳健性的技术。除了可靠的自动驾驶技术带来的社会优势之外，该项目还将把研究结果纳入课程，参与者将指导本科生，并为更广泛地参与计算做出贡献，特别是寻求让洛杉矶中南部克伦肖地区代表性不足的群体的学生接触到计算机。和蒙特贝罗联合学区的学生们一起讨论计算机方面的激动人心的话题。与通用汽车的合作将简化技术转移的道路，并将使博士生接触到与汽车系统相关的主题。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。

Annually, road accidents contribute to approximately 1.35 million fatalities globally. Automated driving and driver assist technology can greatly increase vehicular safety. However, for these technologies, ensuring dependability over a broad set of unusual traffic situations remains a key challenge. To be socially acceptable, these technologies must match or exceed human driving safety levels (e.g. 100 million miles between fatalities). Today, each autonomous vehicle relies on its own sensors to perceive the environment and make independent driving decisions. Unfortunately, these sensors rely on line-of-sight perception and a vehicle's view can be blocked by other vehicles. With advanced wireless communication technologies such as 5G and Qualcomm's Cellular-V2X Technologies, vehicles will be able to share their sensor data with other vehicles, either directly or using roadside infrastructure, so that vehicles can effectively see through obstacles, a capability we call network-enabled cooperative perception. While network-enabled cooperative perception is a compelling technology, the network remains a fundamental bottleneck. Today's vehicular communication technologies can, in practice, achieve about 6-10 Mbps, but advanced vehicular sensors generate hundreds of Mbps of raw data. This project seeks to resolve the tension between the richness of the raw sensor data and the network bottleneck, while scaling network-enabled cooperative perception to extremely dense traffic situations with multi-modal traffic such as pedestrians, bicycles, three-wheelers, trucks, cars etc. in which not all participants may be sensor-equipped. The project will develop abstractions, algorithms, and tools for network-enabled cooperative perception at scale, building upon an abstraction called a glimpse, which is a processed representation of a part of a vehicle's sensor view. Glimpses can represent individual objects within the view, or a grid in 3D space, and can be represented at different granularities that trade-off bandwidth for detail. Given this abstraction, the project will develop: methods to determine, in a complex and highly dynamic traffic setting, which glimpse representations are needed for which vehicles and by when; scheduling algorithms to coordinate the transmission of these glimpses to vehicles while respecting channel capacity constraints; methods to train machine learning models to make control decisions using glimpse-enhanced composite views; and techniques to ensure robustness to glimpse poisoning. Beyond the societal advantages resulting from reliable autonomous driving technology as enabled by network-enabled cooperative perception, the project will incorporate the results of the research into curricula, participants will mentor undergraduates and contribute to efforts to broader participation in computing, specifically seeking to expose students from under-represented groups in the Crenshaw area of South-Central Los Angeles, and students in the Montebello Unified School district to exciting topics in computing. Collaboration with General Motors will ease the path towards technology transfer and will expose PhD students to topics relevant to automotive systems.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.