基于路段-路口联合建模的对偶图卷积网络的行程时间估计方法  

Travel Time Estimation Method based on Dual Graph Convolutional Networks via Joint Modeling of Road Segments and Intersections

作　　者：金广垠 沙恒宇 张金雷 黄金才[1] JIN Guangyin;SHA Hengyu;ZHANG Jinlei;HUANG Jincai(School of Systems Engineering,National University of Defense Technology,Changsha 410005,China;School of Systems Science,Beijing Jiaotong University,Beijing 100089,China)

机构地区：[1]国防科技大学系统工程学院,长沙410005 [2]北京交通大学系统科学学院,北京100089

出　　处：《地球信息科学学报》2023年第7期1500-1513,共14页Journal of Geo-information Science

基　　金：国家自然科学基金项目(62073333、72201029、72288101)。

摘　　要：估计给定路径的行程时间在许多城市交通系统中起着重要作用,例如导航、路线规划和拼车等。然而,现有的大多数工作都侧重于对路段或交叉路口进行单独建模,这并不能准确估计行驶时间,因为交叉路口和路段作为路径的基本要素不仅各自包含多样化的空间属性和时间动态,而且它们之间还具备较强的耦合相关性。为了解决上述问题,本论文提出了一种新颖的端到端深度学习框架,即面向行程时间估计的对偶图卷积网络(DGCN-TTE)来对交叉路口和路段进行联合建模。具体来说,这个模型采用对偶图卷积方法来捕获路口和路段的复杂关系,其中构建节点图来刻画路口之间的相关性,构建边图来表征路段之间的交互特征。为了捕捉空间和时间特征的联合关系,模型中还引入了一种在捕捉时间依赖性的同时结合了从多个邻域范围内整合多尺度空间关系的时空学习模块。本论文通过对3个真实世界的轨迹数据集上的充分的实验来评估提出的DGCN-TTE模型,结果表明该模型显著优于现有的方法,评估指标相比于次优方法最多可以获得超过10%的提升。Estimating travel time for a given route plays an important role in many urban transportation systems,such as navigation and route planning.However,most of the existing works focus on modeling road segments or intersections individually,which cannot accurately estimate travel time,because intersections and road segments,as the basic elements of paths,not only contain diverse spatial attributes and temporal dynamics,but also have strong coupling correlation between them.To address the above problems,this paper proposes a novel end-toend deep learning framework,Dual Graph Convolutional Network for Travel Time Estimation(DGCN-TTE),to jointly model intersections and road segments.Specifically,we adopt a dual graph convolution method to capture the complex relationship between intersections and road segments,in which a node-wise graph is constructed to characterize the correlation between intersections,and an edge-wise graph is constructed to characterize the interaction features between road segments.To capture the joint relationship of spatial and temporal features,we also introduce a spatiotemporal learning module that integrates multi-scale spatial relationships from multiple neighborhood scales while capturing temporal dependencies.This paper evaluates the proposed DGCN-TTE model through experiments on three real-world trajectory datasets,the results show that the proposed model significantly outperforms existing methods,and the evaluation metrics can achieve more than 10%improvement compared with the suboptimal method.

关 键 词：行程时间估计 图卷积网络 时空建模 道路建模 智慧交通 深度学习 神经网络 

分 类 号：U491[交通运输工程—交通运输规划与管理]