Zero-Shot Cellular Trajectory Map Matching

Weijie Shi, Yue Cui, Hao Chen, Jiaming Li, Mengze Li, Jia Zhu, Jiajie Xu, and Xiaofang Zhou · 2025

Read Zero-Shot Cellular Trajectory Map Matching by Weijie Shi, Yue Cui, Hao Chen, Jiaming Li, Mengze Li, Jia Zhu, Jiajie Xu, and Xiaofang Zhou (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-06674.txt.

Introduction

Weijie Shi and co-authors tackle map matching for phone location data in this 2025 paper. Cellular records give rough positions as a user moves. Matching that trail to the right roads is a first step for navigation and route planning. Older methods learn links between cell towers and roads from local data, so they fail in new areas.

The team aims for zero-shot matching, with no fresh training in the target area. A pixel-based tool first adjusts the trail using shared map knowledge. A mixture model sorts scenes so the right expert handles each one. A further module tracks order and position error. Road-level path finding then follows the cleaned trail.

Catalog record

CREATORWeijie Shi, Yue Cui, Hao Chen, Jiaming Li, Mengze Li, Jia Zhu, Jiajie Xu, and Xiaofang Zhou
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-06674

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Zero-Shot Cellular Trajectory Map Matching Weijie Shi, Yue Cui, Hao Chen, Jiaming Li, Mengze Li, Jia Zhu, Jiajie Xu, and Xiaofang Zhou Abstract Cellular Trajectory Map-Matching (CTMM) aims to align cellular location sequences to road networks, which is a necessary preprocessing in location-based services on web platforms like Google Maps, including navigation and route optimization. Current approaches mainly rely on ID-based features and region-specific data to learn correlations between cell towers and roads, limiting their adaptability to unexplored areas. To enable high-accuracy CTMM without additional training in target regions, Zero-shot CTMM requires to extract not only region-adaptive features, but also sequential and location uncertainty to alleviate positioning errors in cellular data. In this paper, we propose a pixel-based trajectory calibration assistant for zero-shot CTMM, which takes advantage of transferable geospatial knowledge to calibrate pixelated trajectory, and then guide the path-finding process at the road network level. To enhance knowledge sharing across similar regions, a Gaussian mixture model is incorporated into VAE, enabling the identification of scenario-adaptive experts through soft clustering. To mitigate high positioning errors, a spatial-temporal awareness module is designed to capture sequential features and location uncertainty, thereby facilitating the inference of approximate user positions. Finally, a constrained path-finding algorithm is employed to reconstruct the road ID sequence, ensuring topological validity within the road network. This process is guided by the calibrated trajectory while optimizing for the shortest feasible path, thus minimizing unnecessary detours. Extensive experiments demonstrate that our model outperforms existing methods in zero-shot CTMM by 16.8%. I Introduction Cellular trajectory data refers to location sequences recorded by cell towers associated with mobile phones, continuously and passively collected on a massive scale of rich information about people’s movements. The widespread of large-scale cellular trajectory datasets has made them indispensable for various location-based applications on web platforms like Google Maps [ 1 ] and Waze [ 2 ] , including navigation [ 3 ] , traffic scheduling [ 4 ] , route optimization [ 5 , 6 ] , and human mobility analysis [ 7 , 8 , 9 ] . To prepare these data for downstream tasks, a key preprocessing step is Cellular Trajectory Map-Matching (CTMM), which aligns the cellular trajectories onto road networks, restoring the paths traveled by users. Unfortunately, compared to high-precision GPS trajectories (with errors of 1-50 meters), cellular trajectories have much higher positioning errors (100-3000 meters), making traditional methods [ 10 , 11 , 12 ] struggle to accurately restore user location based on explicit features and hand-designed heuristics. To mitigate the effects of high location uncertainty, recent learning-based methods like sequence-to-sequence models [ 13 , 14 ] and neuralized HMMs [ 15 ] , use ID-based features to model implicit trajectory-road associations. Nonetheless, ID-based features heavily rely on statistical patterns between specific cell towers and roads within the training data, leading to mismatches in unobserved regions. While annotation data is extremely scarce due to privacy policies, there is a pressing need for zero-shot CTMM capable of generalizing beyond region-specific data. To enable zero-shot CTMM, it is crucial to perform map-matching on a shared geospatial representation. Inspired by TrjSR [ 16 ] , we employ a pixel-based grid map to depict both trajectories and road networks within a shared space. In such way, cellular trajectories and road networks can be transformed into calibrated trajectories at the pixel-level, facilitating accurate user localization. To obtain topologically valid road ID sequences, a path-finding algorithm can be guided by these calibrated trajectories. However, achieving both high accuracy and transferability is not trivial, it presents several significant challenges: • How to enhance the region-adaptive ability. Moving patterns can vary greatly across different regions, in aspects like road network structure, road density, and placement of cell towers. To adapt to these variations in a zero-shot setting, it would be beneficial to customize experts with region-specific knowledge and share knowledge across similar regions, so that the model can quickly adapt to new regions by leveraging relevant knowledge from similar regions it has encountered before. • How to mitigate high location uncertainty. The extensive coverage area of cell towers results in a single tower potentially associating with multiple roads, significantly increasing location uncertainty. To address this, explicitly predicting uncertainty-related characteristics between cell towers and user location allows the model to approximate likely paths. Additionally, extracting temporal features like direction also helps eliminate improbable positions, further refining location estimates. • How to robustly convert calibrated trajectory to road-level path. The ultimate output required for CTMM is a sequence of road IDs, this necessitating the seamless integration of a path-finding algorithm to transform calibrated trajectory from pixel-level into road-level paths. Such a path-finding algorithm should utilize calibrated trajectories as guidance while considering topological accessibility, turn constraints at junctions, and shortest-path preferences to avoid matching detours. In this paper, we propose ZSMM, a two-stage framework for zero-shot map-matching of cellular trajectories. It first calibrates cellular trajectory at the shared pixel space, which then guides the search for a resulting path at the road level. Specifically, ZSMM first pixelates the cellular trajectory and road network into image formats, and then a variational autoencoder (VAE) backbone converts them into a calibrated trajectory image. To share knowledge among similar regions, a Gaussian mixture model (GMM) is integrated into the VAE, where each component is an expert tailored to a type of map-matching scenario. By maximizing the evidence lower bound via stochastic gradient variational Bayes (SGVB), distinctive regions can be separated into different GMM components, enabling dynamic selection of region-adaptive experts. To overcome high location uncertainty, a spatial-aware network empower VAE by explicitly charactering uncertainty-related factors like distance and variance, and a temporal-aware network captures sequential features to eliminate improbable paths. Finally, to ensure topological validity at the road level, a constrained path-finding algorithm is designed to search the optimal path under trajectory calibration constraints while minimizing path length to avoid detours. Extensive experiments demonstrate state-of-the-art zero-shot performance on real-world and synthetic datasets, proving ZSMM’s effectiveness and transferability. The main contributions are summarized as follows: • We are the first to propose a pixel-based trajectory calibration assistant for zero-shot CTMM, followed by constraint path-finding to result in road sequence. Once trained, the model can effectively transfer across unseen regions without any extra training. • We incorporate a GMM into VAE to enhance region-adaptive ability by knowledge sharing among similar regions while customizing for distinct areas. • We design a spatial-aware network to explicitly characterize location uncertainty, and a temporal-aware network to capture sequential features, jointly mitigating the high uncertainty inherent in cellular trajectories. • We design a constrained path-finding algorithm that takes the calibrated trajectory as guidance while minimizing path length to avoid detours. II RELATED WORK Map-matching is a critical preprocessing step for location-based web applications, aiming to align trajectory points with road networks to identify the path traveled by a vehicle or user. Compared to GPS trajectories, cellular trajectories have larger-scale data, as cell towers passively collect user mobility information at all times [ 17 ] . Unfortunately, they also have much higher positioning errors, posing tremendous challenges for map-matching [ 18 ] . A variety of map-matching algorithms have been proposed for cellular trajectories, which can be categorized into traditional methods and learning-based methods [ 19

Read the complete text