Enhance 3D Point Cloud Accuracy Through Supervised Machine Learning for Automated Rolling Stock Maintenance: A Railway Sector Case Study

This paper presents findings of a case study conducted to introduce industrial robots into automatic train coupler inspection of Siemens Class 380 rolling-stock. The targets are localized by coalescing RGB and time of flight (ToF) sensor data. The study examines several supervised machine learning techniques to improve the overall accuracy of 3D point clouds. A cost factor which reflects root mean square, mean absolute error and coefficient of determination is defined to evaluate the performance of the learning algorithms. The best-suited models are further validated using simulation data and selected to include in overall robotic sensing system.