@phdthesis{woltche2026_116612,
  title = {Map matching methods for assigning vehicle positions to road networks},
  author = {Wöltche, Adrian},
  year = {2026},
  publisher = {University of Bamberg Press},
  address = {Bamberg},
  series = {Schriften aus der Fakultät Wirtschaftsinformatik und Angewandte Informatik der Otto-Friedrich-Universität Bamberg},
  isbn = {978-3-98989-134-0},
  doi = {10.20378/irb-116612},
  urn = {urn:nbn:de:bvb:473-irb-116612x},
  url = {https://fis.uni-bamberg.de/handle/uniba/116612},
  language = {eng},
  abstract = {Map matching is a widely used technology for resolving discrepancies between sequences of geocoordinates from movements and road networks. During movement, geocoordinates can be measured with the help of global navigation satellite systems. However, due to the nature of these systems, the measured geopositions contain considerable noise. Consequently, discrepancies generally occur between the measured positions and the underlying roads on which the movement originally occurred.

Map matching is a combination of defined models, algorithms, and metrics for correcting these discrepancies. The goal of map matching is to find routes within the road network that most accurately represent the recorded movements. The results of map matching exactly overlap with the road network, are free from the original noise, and are corrected concerning impossible detours and gaps from the measured movements. These results facilitate further analyses in the domain of mobility research.

In this thesis, we present a novel map matching method based on Markov decision processes and reinforcement learning algorithms. Our novel method introduces various advances in multiple aspects of the map matching process pipeline for increasing the accuracy of our results in comparison to existing solutions. Our process includes an improved trajectory simplification preprocessing step for improving matching performance both in terms of accuracy and speed, a novel candidate adoption feature that enables the stochastic process to dynamically resolve outliers and noise clusters, a new metric that evaluates direction changes in routes for reducing detours and round trips in the results, and finally, a new comparison algorithm that can evaluate results from different map matching solutions with varying numerical coordinate precision.

Our novel map matching solution is provided as ready-to-use open source software, and we compare our solution with multiple third-party open source solutions in terms of computational speed, resource consumption, and accuracy of the results. For achieving high performance, our open source reference implementation contains several specific data structures, algorithmic customizations, and software design decisions, that we explain in context.

For our extensive benchmark, we use multiple existing map matching data sets, with a mixture of ground truth and hand-corrected solutions. Furthermore, we introduce and analyze a novel open source process for generating artificial ground truth map matching data. Our evaluation demonstrates that our novel map matching method is more robust to noise and outliers than existing methods and that it outperforms them in terms of accuracy and computational performance.

Our new map matching approach contains several key concepts that can be individually transferred to other solutions. Additionally, our new approach generates further ideas for future research, some of which we present in this work.},
  keywords = {map matching, open source, benchmark, geographic information science, markov decision process, reinforcement learning},
  school = {Otto-Friedrich-Universität Bamberg}
}
