定位与通信受限的网联协同感知算法  

V2X-Enabled Cooperative Perception with Localization and Communication Constraints

作　　者：毛瑞清 贾宇宽 孙宇璇 周盛[1] 牛志升[1] MAO Ruiqing;JIA Yukuan;SUN Yuxuan;ZHOU Sheng;NIU Zhisheng(Department of Electronic Engineering,Tsinghua University,Beijing 100084;School of Electronic and Information Engineering,Beijing Jiao-tong University,Beijing 100044)

机构地区：[1]清华大学电子工程系,北京100084 [2]北京交通大学电子信息工程学院,北京100044

出　　处：《模式识别与人工智能》2023年第11期1019-1028,共10页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.62341108,62221001,62022049,62111530197);清华大学-丰田自动驾驶汽车人工智能技术联合研究中心资助~。

摘　　要：随着车联网通信技术的不断发展,协同感知的网联自动驾驶成为未来智慧交通系统中重要的组成部分,可有效解决传统单车智能在感知和计算能力上的固有局限.然而,现有的协同感知算法大多依赖准确的定位信息进行多视角融合,往往忽略通信带宽与通信时延的约束.因此,文中提出面向定位与通信受限条件下的特征层级协同感知算法,在不依赖准确位置与位姿信息的情况下实现不同视角信息的匹配,同时对通信时延具有一定的鲁棒性,能根据信道状态动态调整通信数据量.文中算法结合传统两阶段感知范式与深度度量学习,通过传输区域特征图进行跨视角信息匹配,降低定位误差和通信时延带来的影响.同时,通过车联网通信传输的区域特征图数量可以实时在线调整通信数据量,满足不同信道条件的约束.在协同感知数据集上的实验表明,文中算法在各场景下均表现出显著的协作增益,在一定通信时延下感知精度不显著衰减,可有效降低所需传输的数据量.With the continuous development of vehicle-to-everything network,cooperative perception enabled connected autonomous driving becomes an important component in future intelligent transportation systems.It effectively addresses inherent limitations of traditional stand-alone intelligence in perception and computing capabilities.However,most existing cooperative perception algorithms rely on accurate positioning information for data fusion,ignoring constraints of communication bandwidth and commu-nication delay.In this paper,a feature-level cooperative perception algorithm for localization and communication-constrained conditions is proposed.The matching of different perspective information is achieved without relying on accurate positions and poses,while the robustness of the proposed algorithm to communication delay is maintained and the amount of communication data is dynamically adjusted according to the channel state.The traditional two-stage perception paradigm is combined with deep metric learning,utilizing regional feature maps for cross-perspective information matching to overcome the impact of localization errors and communication delays.Moreover,the number of regional feature maps transmitted through V2X communication can be dynamically adjusted in real-time to adapt to different channel conditions,and thus the amount of communication data is changed.Experimental results show that the proposed algorithm exhibits significant cooperative gains in various scenarios,maintains perception accuracy under certain communication delays,and effectively reduces the required amount of transmitted data.

关 键 词：协同感知 车联网通信 网联自动驾驶 度量学习 

分 类 号：TN92[电子电信—通信与信息系统]