Railway reconstruction from 3D point cloud using Deep Learning and Parametric Modeling

Abstract. Railway infrastructure is crucial for transporting goods and passengers, making its maintenance and reconstruction vital for safety and reliability. Traditional methods reliant on manual surveys are time-consuming and prone to inaccuracies. Although 3D point cloud data provides detailed representations of railway environments, its unstructured nature complicates processing and modeling. This paper presents a methodology that combines deep learning with parametric modeling to reconstruct railway environments from 3D point cloud data, focusing on key components such as rails, catenary wires and poles. The results are represented in a standardized CityJSON format, in compliance with the Transportation module of CityGML 3.0, and textured to create photo-realistic 3D railway models. The proposed approach uses the KPConv (Kernel Point Convolution) architecture for semantic segmentation to classify railway components. The model is trained on Rail3D dataset and achieved a mean Intersection-over-Union (mIoU) of 84%. Instance segmentation of catenary poles is performed using Label Connected Components (LCC) algorithm, followed by a second-level classification through template matching using Fast Global Registration (FGR) and Iterative Closest Point (ICP). Rail reconstruction combines Region Growing and H-DBSCAN algorithms for clustering, vectorization for linear geometry extraction, and extension to ensure continuity despite gaps or noise in the data. Catenary poles are reconstructed using parametric models, taking as input a scale factor and a rotation matrix calculated from the extracted height and azimuth. Wires are added accordingly to connect the reconstructed poles. The methodology was validated on Belgian railway data, producing accurate, interoperable and photo-realistic 3D models suitable for digital twin integration, infrastructure monitoring and urban simulations.