This thesis presents a methodology for inferring the operational track characteristics of a traveled railway path with only GPS trajectory data from a train as input. The approach enriches the OpenStreetMap (OSM) railway network with detailed track parameters that are crucial for advanced engineering applications but are not available in public map sources. The proposed method first reconstructs a train’s path from noisy, raw GPS trajectories using a map-matching algorithm. Subsequently, a machine learning model infers key rail parameters along the identified path. These include track cant and horizontal geometric track alignment elements, which define the track layout as a sequence of straights and curve types, such as circular and transitional curves.

To bridge this data gap, this work develops and implements an integrated two-stage computational pipeline. A crucial prerequisite for parameter inference is accurately determining the path of the train, the first stage introduces a robust and highly efficient map matching algorithm, the Fast Railway Matcher. This flexible method, based on a Hidden Markov Model (HMM) framework and specifically adapted to railway characteristics, effectively aligns noisy GPS trajectories with the OSM railway map across multiple operational areas and varying GPS sampling rates. The second stage employs a deep learning framework where Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models are implemented within a multi-task learning setup. Using a dataset of more than 3000 km of German railway infrastructure as ground-truth, this setup utilizes a shared network backbone to simultaneously perform regression for track cant and classification for the horizontal geometric track alignment elements, with a loss function that dynamically balances both tasks.

Experimental results validate the effectiveness of the proposed pipeline. On a ground-truth trajectory between Mönchengladbach and Aachen, the Fast Railway Matcher achieves an F1-score of over 0.93 and demonstrates a runtime speed-up of approximately 30-fold on a 30-day operational dataset from the Rhein-Ruhr-Express compared to established literature methods. The deep learning-based rail parameter inference models were evaluated on a test split of the ground-truth dataset. Since ground-truth curvature is also unavailable in OSM, the evaluation critically compares using interpolated ground-truth curvature against a practical spline-fitting estimation method as an input feature. With interpolated ground-truth curvature, the LSTM model achieves near-perfect results. When using the estimated curvature, the LSTM model’s performance remains strong, achieving a mean absolute error of 0.014 meters for cant prediction and a classification accuracy of approximately 90%. This highlights the accuracy of the curvature estimation as the most critical input feature and the primary bottleneck of the inference task. This research provides a comprehensive methodology to transform readily available GPS data into actionable track information, enabling significant advancements in the design, analysis, and maintenance of trains and infrastructure.