University of Twente Student Theses


Machine learning classification of objects from RBG images and point clouds obtained from a MLS system in railroad environment

Loganathan, Athithya Seethalakshmi (2020) Machine learning classification of objects from RBG images and point clouds obtained from a MLS system in railroad environment.

Link to full-text:
(only accessible for UT students and staff)
Abstract:The data fusion of RGB images and 3d point clouds captured from a Mobile Laser Scanning(MLS) platform is gaining more research interest recently due to its application in various fields such as railway infrastructure management, road traffic infrastructure maintenance, and autonomous vehicle navigation. Especially in the railway industry, periodical surveys are undertaken by mounting the MLS platform to the rails. And the data captured from the surveys are further post-processed to obtain a railway infrastructure model. However, the creation of such infrastructure models is time-consuming and involves a lot of manual labor, as understanding and identifying objects in the scene is quite difficult. The overall time taken to create such models can be drastically reduced by introducing a machine-learning algorithm to perform the classification task. This research proposes a novel framework that leverages both RGB images and 3D point clouds for efficient inventory mapping. The research segments are – object detection and classification, Character recognition from detected objects, and Positioning the detected objects in the point clouds. For this study, two objects of interest are selected.; they are - kilometer markers and Signals. Training samples for these objects are created from the RGB images, and a machine learning classifier is trained using these samples for object detection. As a result, the object's location in the RGB images is identified. Since the kilometer marker contains digit values in it, the detected kilometer marker images are further processed for character recognition. The task of character recognition is performed using a deep learning model that is being trained to recognize digits from a natural environment. The recognized kilometer values attribute to the semantic information of the objects. Furthermore, to obtain the 3d geometry, the objects detected in the image are reprojected to 3d point clouds using internal spatial parameters as they are captured from the same platform. With some processing, the geometry and its subsequent point cloud structure of the objects are obtained and thus locating the object in the real world accurately. The different classifiers used in this study are tested for their performance in classifying the objects, and their accuracy is assessed. Among them, the multi-class SVM with RBF kernel is selected for further object detection. It produces an overall accuracy of 96%. Similarly, the character recognition model is evaluated for its performance, and the accuracy is assessed. The algorithms are implemented in python, and the efficiency of the framework is evaluated. --- For access to the full thesis, please contact the ITC Library (
Item Type:Essay (Master)
Faculty:ITC: Faculty of Geo-information Science and Earth Observation
Programme:Geoinformation Science and Earth Observation MSc (75014)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page