顾及路网约束的深度地图匹配方法  

Deep learning-based map matching considering road network constraints

作　　者：钟青岑 吴晨昊 向隆刚[1,3] 姚鹏 ZHONG Qingcen;WU Chenhao;XIANG Longgang;YAO Peng(State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China;Fujian Expressway Science&Technology Innovation Research Institute Co.,Ltd.,Fuzhou 350001,China;Hubei Luojia Laboratory,Wuhan 430079,China;Guangxi Taiwei Information Technology Co.,Ltd.,Guilin 541100,China)

机构地区：[1]武汉大学测绘遥感信息工程国家重点实验室,湖北武汉430079 [2]福建省高速公路科技创新研究院有限公司,福建福州350001 [3]湖北珞珈实验室,湖北武汉430079 [4]广西泰绘信息科技有限公司,广西桂林541100

出　　处：《测绘通报》2024年第6期96-102,133,共8页Bulletin of Surveying and Mapping

基　　金：湖北省珞珈试验室专项基金(220100010);广西JMRH发展专项项目(202203)。

摘　　要：在低频或非均匀采样条件下,已有的地图匹配算法存在匹配精度不高或效率较低的问题。本文提出了一种顾及路网约束的深度地图匹配方法(RNCMM)。该方法首先利用Seq2Seq框架将低频轨迹点序列端到端地映射为高频路段序列;其次根据道路与轨迹点间的距离、方位差构建细粒度约束掩模层,有利于缓解轨迹网格表示的局限性,提高匹配精度;然后引入注意力机制和多任务学习机制,挖掘轨迹点间的时空关联性,并进行路段与方向的联合预测;最后在Porto出租车轨迹数据集和OSM路网上进行试验。结果表明,相较于传统的隐马尔可夫模型(HMM)算法,本文方法可以有效地提高低频浮动车轨迹的匹配精确度和效率。In low-frequency or non-uniform sampling conditions,existing map matching algorithms have problems of low matching accuracy or low efficiency.In this paper,we propose a road network constrained map matching model based on deep learning(RNCMM).Firstly,Seq2Seq framework is used to map the low frequency track point sequence to the high frequency road segment sequence from end to end.Secondly,a fine-grained constraint mask layer is constructed according to the distance and azimuth difference between the road and the trajectory point,which is conducive to alleviating the limitations of the trajectory grid representation and improving the matching accuracy.Then,attention mechanism and multi-task learning mechanism are introduced to mine the spatiotemporal correlation between trajectory points and perform joint prediction of road segments and directions.Finally,experiments are conducted on the Porto taxi trajectory dataset and OSM road network.The results show that compared to traditional hidden Markov model(HMM),the proposed algorithm can effectively improve the matching accuracy and efficiency of low-frequency floating car trajectories.

关 键 词：地图匹配 深度学习 序列到序列模型 GRU 多任务学习 注意力机制 

分 类 号：P28[天文地球—地图制图学与地理信息工程]