城市复杂环境智能驾驶基础理论与关键技术研究

It is an urgent demand to provide remote information service and driving safety service for intelligent driving systems. The existing driving systems are mainly based on vehicle-to-vehicle and vehicle-to-infrastructure/cloud communication, with vehicle-to-road communication as auxiliary. However, this communication scheme may result in bad practicality, or cannot meet the performance requirements, such as real-time, reliable and privacy preserving, of two services for intelligent driving. However, edge computing can provide nearby edge-intelligent services and satisfy the requirements of real-time, reliable and privacy preserving by merging networking, computation, storage and other functions together. Hence, aiming at effective and secure intelligent transportation services, this project tries to solve the requirements and challenges of these two kinds of services with the consideration of complex urban traffic environment. We will adopt the edge computing technique into intelligent driving applications. The three main scientific problems needed to be solved are how to provide effective and reliable integration transmission mechanism in complex city environment, how to implement the vehicle-road-edge collaboration mechanism and how to protect users' privacy for intelligent driving. As a result, the project mainly addresses four issues of intelligent driving system as follows: the modeling of complex city environment and intelligent driving, the intelligent mechanisms of sensing and processing on the vehicular terminals, the reliable and real-time vehicle -edge-cloud/infrastructure interaction mechanism and privacy preservation orienting to intelligent decision on the cloud platform. Finally, we will attempt to develop an intelligent driving demonstration system in the complex city environment, and facilitate the performance analysis and verification for the designing protocols and algorithms.

远程信息服务和行车安全服务是智能驾驶的两类重要需求，现有的以车车/车网通信为主、车路通信为辅的驾驶体系实用性差，无法为智能驾驶提供满足实时可靠与隐私保护等性能的两类服务。而边缘计算通过融合网络、计算、存储等功能，就近提供边缘智能服务，可满足实时可靠与隐私保护等需求。故本项目以城市道路/交通环境的复杂性为研究立足点，以提供高效、安全的智能驾驶服务为主要目标，综合分析两大服务面临的需求与挑战，将边缘计算技术引入智能驾驶应用中，紧紧围绕城市复杂环境下高效可靠的一体化传输机制、车路边协同机理、面向智能驾驶的隐私保护机制这三个科学问题，重点研究满足不同服务质量与需求的智能驾驶应用的基础理论和关键技术，具体包括城市复杂环境与智能驾驶的建模、车载终端侧智能感知与处理、车边云交互实时可靠保障和云平台侧隐私保护的智能决策，最终开发城市复杂环境下智能驾驶示范与验证系统，对设计的协议和算法进行验证分析。

本项目以远程信息服务和行车安全服务为应用背景，以城市环境的复杂性和动态性为研究立足点，综合分析智能驾驶应用所面临的需求与挑战，研究满足不同服务质量与需求的智能驾驶保障基础理论与关键技术。在该项目的资助下，课题组在国际著名会议和期刊上发表高水平论文93篇，其中CCF推荐A类论文48篇（包括INFOCOM 10篇、ToN 8篇、TMC 7篇等），B类论文22篇（包括ComNet 6篇、TSC 6篇等）。申请专利14项，授权1项。主要学术贡献包括：（1）为实现对智能驾驶数据的高效分析与预测，课题组设计了一个基于车边协同的分布式点对点（P2P）模型训练架构。在每轮训练中,根据系统资源开销与网络链路状态，利用深度强化学习算法动态构造一个P2P通信拓扑，以此进行车载终端之间的模型传输与聚合，提高数据分析效率约11%，降低通信开销约30%。（2）将车载终端感知任务和数据迁移至边缘服务器进行处理时，课题组提出了结合服务部署的任务迁移与数据融合算法，与基准算法相比，所提算法可以提高系统吞吐量56%-69%。（3）针对车边云交互实时可靠保障研究，课题组首次研究了能容忍多服务器故障的健壮性任务卸载问题，提出了在任务到达时对其进行卸载的在线原始对偶算法，该算法能够很好地应对边缘网络动态性，保障了服务响应的低延迟。（4）针对车辆可能不信任云平台或者边缘服务器从而导致其不愿提供隐私数据的问题，课题组提出了可实现隐私保护的车辆选择算法，该算法能够以较小的成本代价保证系统性能，此外算法在隐私保护方面的运行时间也要远小于同态加密算法和乱码电路算法。项目执行期内，组织网络大数据与信息安全研讨会1次，培养国家优青1名、博士/硕士研究生23名，申请人后续也获批了基金委重点项目，获安徽省自然科学二等奖。通过该项目的实施，深入挖掘了车、路、边等因素对智能驾驶系统的影响，增强了驾驶的智能化程度，有望提高我国城市智慧化建设的水平。