Trans2Vec:耦合车辆移动模式的大规模城市路网表示学习框架  

Trans2Vec:a representation learning framework for large⁃scale urban road networks coupled with vehicle movement patterns

作　　者：邓敏[1] 储国威 谌恺祺 石岩[1] DENG Min;CHU Guowei;CHEN Kaiqi;SHI Yan(Center South University,Changsha 410083,China)

机构地区：[1]中南大学,湖南长沙410083

出　　处：《测绘通报》2023年第3期74-74,75-78,共5页Bulletin of Surveying and Mapping

基　　金：国家自然科学基金(42171459,41730105)。

摘　　要：高效率实现城市路网信息的完备向量化表征,是嵌入各类深度学习模型以智能化解决下游交通任务的重要数据工程。现有表示学习方法难以有效耦合城市路网固有的拓扑结构信息与承载的转移模式信息,无法精确匹配下游交通任务的特征需求,且在面对大规模城市路网时,存在算法资源占用较高的现实问题。针对该问题,本文发展了一种多视图游走的路段表征向量学习方法,以车辆移动模式为核心,融合拓扑结构进行异质游走,实现大规模路网完备、高效的信息表达;以深圳市为试验区域,通过全量城市路网与实际车辆行程数据验证了该方法的有效性与先进性。Efficient learning of representation of urban road network information is an important project for embedding various deep learning models to intelligently solve downstream traffic tasks.However,the existing learning framework are difficult to effectively couple the topology information of the urban road network and the transfer mode information carried,and cannot accurately match the characteristic requirements of downstream traffic tasks.In this reason,this paper develops a multi⁃view roaming segment representation vector learning framework,which takes the vehicle movement pattern as the core,integrates the topology structure for heterogeneous walk,and realizes the complete and efficient information embedding of large⁃scale road network.In the experimental part,taking Shenzhen as the case area,the effectiveness and advanced nature of the method in this paper are verified through the full urban road network and actual vehicle travel data.

关 键 词：Trans2Vec 城市路网 车辆移动模式 表示学习 

分 类 号：P208[天文地球—地图制图学与地理信息工程]