情景路网约束下基于序列到序列的轨迹恢复方法

PDF下载HTML

Title:

A seq2seq based trajectory recovery method under the constraint of scenario road network

作者:

孟祥福, 葛檄文, 杨雨卓

辽宁工程技术大学 电子与信息工程学院, 辽宁 葫芦岛 125105

Author(s):

MENG Xiangfu, GE Xiwen, YANG Yuzhuo

School of Electronic and Information Engineering, Liaoning Technical University, Huludao 125105, China

关键词:

路网约束; 出行模式; 稀疏轨迹; 轨迹恢复; 序列到序列; 情景路网; 时空特征提取; 路段推理

Keywords:

road network constraint; travel pattern; sparse trajectory; trajectory recovery; sequence-to-sequence; scenario road network; spatiotemporal feature extraction; road segment inference

分类号:

TP183

DOI:

10.11992/tis.202506009

摘要:

在真实世界中，轨迹数据日益激增却又稀疏，轨迹恢复对下游轨迹挖掘任务具有重大意义。路网约束下端到端的轨迹恢复方法大多通过静态路网和轨迹坐标描述轨迹并嵌入向量，既忽视道路动态流量变化，又未考虑驾驶者出行模式与区域、路段功能的关联，而这些特征难以从稀疏轨迹时空特征和道路建设语义特征中提取。为此，本文提出一种情景路网约束下基于序列到序列的轨迹恢复方法(road network-constrained scenario-based trajectory recovery, CSTrajRec)，该方法将轨迹输入傅里叶编码层提取其时空特征，强化对整体模式的表达。在编码器中，用路段所在位置的兴趣点文本、道路种类和时间段内的区域流量建模每个路段，然后用路段描述轨迹序列以反映个体的出行模式和偏好。在解码器中，设计一种新颖的融合距离和道路情景的路段推理模块用于指导轨迹的恢复。通过在3个真实轨迹数据集上构建的不同采样间隔稀疏轨迹的实验，充分验证了本文所提出模型的优越性。

Abstract:

In the real world, trajectory data grows exponentially yet remains sparse, making trajectory recovery crucial for downstream mining tasks. Most existing end-to-end trajectory recovery methods under road network constraints embed static road networks and coordinates into vectors. However, they overlook dynamic road traffic flow and neglect links between drivers’ travel patterns and the functions of regions/road segments, which are hard to extract from sparse trajectory spatio-temporal features and road semantic features. This paper proposes a road network-constrained scenario-based sequence-to-sequence trajectory recovery method (CSTrajRec). It inputs trajectories into a Fourier encoding layer to extract spatio-temporal features and enhance overall pattern representation. In the encoder, each road segment is modeled with text-based points of interest, road type, and regional traffic, which are then used to describe the trajectory sequence for individual travel patterns. In the decoder, a novel segment inference module fusing distance and road scenario information is designed for trajectory recovery. Experiments on three real datasets with sparsely sampled trajectories at varying intervals validate the model’s superiority.