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.

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。