智能网联环境下城市道路多源交通数据补全方法  

Multisource Traffic Data Completion Method for Urban Roads in Intelligent Connected Scenarios

作　　者：王庞伟[1] 何昕泽 张龙[1] 董航瑞 王力[1] 张名芳 WANG Pang-wei;HE Xin-ze;ZHANG Long;DONG Hang-rui;WANG Li;ZHANG Ming-fang(Beijing Key Lab of Urban Intelligent Traffic Control Technology,North China University of Technology,Beijing 100144,China;Beijing Connected and Autonomous Vehicles Technology Co.Ltd.,Beijing 100176,China)

机构地区：[1]北方工业大学城市道路交通智能控制技术北京市重点实验室,北京100144 [2]北京车网科技发展有限公司,北京100176

出　　处：《中国公路学报》2025年第1期281-293,共13页China Journal of Highway and Transport

基　　金：车路一体智能交通全国重点实验室开放基金课题(2024-A001);国家重点研发计划项目(2022YFB4300400);北京市自然科学基金(4212034)。

摘　　要：交通状态补全方法能够为交通管理系统提供完备的全息交通路网信息,为制定城市信控策略,动态均衡交通流提供数据支持。基于智能网联技术实时获取多源交通数据优势,提出一种基于图卷积神经网络的实时交通状态补全方法。首先,构建了一种“端-边-云”信息交互架构的全息交通感知系统,可实现多源交通数据的特征级融合;其次,根据路网拓扑关系构建路网无向图模型,应用异常数据辨识与插补方法对原始数据进行修正构成有效数据集,并根据实际路网时空关系确定补全网络隐藏层权重;然后,通过图卷积交叉口临近关系与交通状态,将原始数据映射至空间维度,从而完成交叉口特征的空间聚类,同时由门控循环单元在时间序列上游走记忆,提取数据时间维度特征,完成状态数据补全计算;最后,在北京市高级别自动驾驶示范区选取典型智能网联交叉口群,对该方法进行实地测试。研究结果表明:长时序数据下,方法有效补全结果与真实值误差不高于10.64%,综合性能较长短期记忆神经网络等现有方法的均方根误差降低17.2%。该补全方法为未来智能网联环境下交通全息感知技术应用提供了理论基础和实现方案。Traffic state completion methods can provide comprehensive holographic traffic network information for traffic management systems,supporting the formulation of urban signal control strategies and dynamic balancing of traffic flow.Leveraging the advantages of real-time acquisition of multi-source traffic data through intelligent and connected technologies,this paper proposes a real-time traffic state completion method based on graph convolutional neural networks.First,a holographic traffic perception system with an“end-edge-cloud”information interaction architecture was constructed,enabling feature-level fusion of multi-source traffic data.Second,an undirected graph model of the road network was built based on the road network topology.Anomaly data identification and interpolation methods were applied to correct the raw data,forming an effective dataset.The hidden layer weights of the completion network were determined according to the spatiotemporal relationships of the actual road network.Third,the spatial features of intersections were clustered by mapping the original data to the spatial dimension through the graph convolutional approach,which incorporates adjacency relationships and the traffic states of intersections.The gate recurrent unit(GRU)was used to traverse the data along the time series,extracting temporal features for state data completion calculations.Finally,field tests were conducted at typical intelligent and connected intersections in the Beijing High-level Automated Driving Demonstration Area.The test results show that for long-term sequence data,the method achieved an error of no more than 10.64%compared to the real values.The overall performance,as measured by the reduction in root mean-squared error(RMSE),was 17.2%lower than that of existing methods such as the long short-term memory(LSTM)neural network.This completion method provides a theoretical foundation and implementation solution for the application of traffic holographic perception technology in future intelligent and connected

关 键 词：交通工程 交通数据补全 图卷积神经网络 长短期记忆神经网络 边缘计算 智能网联交通 

分 类 号：U491.2[交通运输工程—交通运输规划与管理]