Accurate and up-to-date maps of exising rail infrastructure is often not available. Currently, information from the railway scene is surveyed by human operators when it is needed. This is a costly and time consuming operations, furthermore the rail cannot be used during surveying for safety reasons.
Goal of this research was to automate the survey process and deliver fast, accurate maps without disturbing the availability of the rail. This is done by using a train mounted laser scanner and using a modified PointNet++ network to perform the semantic segmentation task. Semantic segmentation assigns a meaningful label to each point within the point cloud.
Based on a leave-one-out-cross-validation using a modified PointNet++ network an mIoU of 71% is obtained. This is a promising result. If repeated measurements over time are taken, this work can also pave the way towards predictive maintenance.