融合图结构学习和轻量级循环建模的地图匹配方法  

Leveraging Graph Structure and Simple Recurrence for Map Matching

作　　者：罗威 刘钰 黄强 武志昊[1] LUO Wei;LIU Yu;HUANG Qiang;WU Zhihao(School of Computer and Information Technology,Beijing Jiaotong University,Beijing 100044;Tencent Inc.,Beijing 100193)

机构地区：[1]北京交通大学计算机与信息技术学院,北京100044 [2]腾讯科技(北京)有限公司,北京100193

出　　处：《北京大学学报（自然科学版）》2024年第6期979-988,共10页Acta Scientiarum Naturalium Universitatis Pekinensis

基　　金：中央高校基本科研业务费;腾讯犀牛鸟大出行专项研究计划资助.

摘　　要：现有的地图匹配方法主要依赖序列到序列模型来捕获轨迹内关联性,忽略路段间、轨迹间以及轨迹与路段间的关联性。同时,现有方法采用的循环神经网络因其固有结构,难以进行高效的并行计算。为了充分利用数据中存在的多种关联性,并提升模型的并行计算能力,提出一种融合图结构学习和轻量级循环建模的地图匹配方法(GMMSR)。通过路网卷积和轨迹图卷积,建模路段之间和轨迹之间的关联性,采用在隐空间对齐路网和轨迹表示的方式,建模轨迹与路段之间的关联性。利用轻量级循环单元实现模型更高效的并行计算。在北京市某区域轨迹路网数据集上的实验结果表明,所提模型较已有基准模型在精度上实现大幅度提升,在效率上相当或更好。Existing solutions for map matching mainly rely on sequence-to-sequence models to capture the correlations within a trajectory while neglecting the correlation between road segments and trajectories as well as trajectory-road correlations.Meanwhile,recurrent neural networks suffer from inherent limitations in conducting computations efficiently in parallel.To fully exploit all the aforementioned correlations and to improve the model parallelism,a Graph-enhanced Map Matching model with Simple Recurrence(GMMSR)is proposed.The model captures the correlations between road segments and trajectories through road network convolution and trajectory graph convolution respectively,and exploits the trajectory-road correlation by aligning road network and trajectory representations in latent space.Moreover,the model utilizes simple recurrent units to achieve more efficient parallel computations.Extensive experiments on a map matching dataset in a subarea of Beijing demonstrate significant improvements in accuracy compared with existing baselines while achieving comparable or better efficiency.

关 键 词：地图匹配 轨迹和路网关联性 图神经网络 轻量级循环单元 

分 类 号：U463.6[机械工程—车辆工程] TP18[交通运输工程—载运工具运用工程]