资源受限下道路网络临近检测技术研究

Proximity detection in road networks is to detect which of the moving clients in road networks are close to each other, which is of great importance to guarantee traffic fluency and safety, ensure auxiliary driving, and realize large-scale automatic driving. Most of existing proximity detection solutions are based on geographical distance, and adopt traditional centralized client-server architecture or distributed peer-to-peer architecture, facing the problems of high communication latency, high communication cost, and long computational time, under the condition of limited communication resources and limited computing resources. Therefore, given that time distance is more valuable to time-aware proximity detection in road networks than geographical distance, this project aims to do research on key technologies of proximity detection in road networks based on time distance in resources-constrained environment, aiming to propose (1) proximity detection architecture based on mobile edge computing, (2) intelligent proximity detection algorithms based on deep learning, and (3) "offline preprocessing - online proximity detection" computational time optimization scheme based on GPU acceleration, so as to achieve the goal of reducing communication latency, communication cost and computational time simultaneously for proximity detection in road networks in resources-constrained environment.

道路网络临近检测是要实时检测出道路网络中哪些移动用户彼此临近，这是保证交通畅通安全、保障辅助驾驶、实现大规模自动驾驶的基础。现有道路网络临近检测方法大多基于地理空间距离采用传统的集中式的client-server架构或分布式的P2P架构，在通信资源和计算资源受限的条件下面临着通信时延高、通信成本高、计算时间长等瓶颈问题。因此，考虑到时间距离比地理空间距离对时间感知的道路网络临近检测更有价值，本项目拟开展资源受限下基于时间距离的道路网络临近检测关键技术研究，拟提出①基于移动边缘计算的道路网络临近检测架构，②基于深度学习的智能临近检测算法，③基于GPU加速的“线下预处理-线上临近检测”计算时间优化机制，以达到在资源受限条件下同时降低道路网络临近检测的通信时延、通信成本、计算时间的目标。

道路网络临近检测对于实现未来大规模自动驾驶具有重要意义。然而，由于动态道路网络中通信资源和计算资源受限，如何实现低成本的临近检测成为一个难题。因此，本项目开展了资源受限下道路网络临近检测技术研究。. 首先，基于移动边缘计算技术，设计了新型临近检测架构，保留了原有临近检测架构的client与server之间的通信机制，在多个边缘云部署多个边缘服务器，克服了传统client-server架构或peer-to-peer架构通信成本高的缺点，显著降低了道路网络临近检测通信时延成本。. 其次，设计了道路网络临近检测算法，在实现基本临近检测的同时，能通过自动调整安全区域半径达到最小化临近检测通信消息数（通信成本之一）的目标；提出了“线下预处理-线上临近检测”的计算机制，并使用多线程计算方法，降低了计算成本。. 再次，利用凸优化方法，提出了降低道路网络临近检测能耗以及通信时延的模型、算法及解决方案，实现最小化道路网络临近检测能耗以及通信时延的目的；基于遗传算法NSGA-II和深度强化学习DDPG算法，提出了联合优化通信时延和能耗的模型、算法及解决方案，达到进一步降低通信的时延成本和能耗成本的目标。. 基于本项目研究，项目组发表论文多篇，授权专利多项，项目成果可以为车联网大规模自动驾驶提供参考。

内容获取失败，请点击重试