Uncovering Railway Infrastructure: AI-Driven Segmentation of Dense Point Clouds

Abstract

This thesis explores the potential of using advanced Artificial Intelli- gence techniques to automate the segmentation of railway point clouds to facilitate creating accurate object inventories. The motivation for this research stems from the need to modernise outdated engineer- ing drawings of railway catenary systems, which have become unre- liable due to temporary fixes and undocumented changes over time. By implementing novel segmentation models, specifically KPConvX and PTv3, this work aims to accurately segment railway infrastructure elements in diverse and complex scenes, including multi-lane tracks, crossings, bridges, tunnels and stations. The research questions fo- cus on comparing these advanced models with previous methods, as- sessing their generalisation capabilities, and extracting specific compo- nents such as section insulators and jumper wires. The contributions of this thesis include the introduction of state-of-the-art segmentation techniques to the railway domain, the segmentation of previously un- addressed objects, and the extension of an existing dataset with la- belled point cloud data from the Netherlands, France and Hungary. Ultimately, this work aims to support predictive maintenance and au- tomated inspections, thereby contributing to the safety and efficiency of railway infrastructure.