Abstract
Railway systems are among the most important transportation routes. Due to their vastness, limited evasion possibilities for trains, and their long braking distances, ensuring the safety of these networks is challenging and can be improved through the application of deep learning. While the localization of objects in the track bed is a common topic, the classification of obstacles is not yet solved robustly for all relevant classes at a high-performance level. This study proposes an architecture tailored to the specific characteristics of track bed images based on a ResNet-50 backbone, leveraging individual characteristics of those images (i.e. high frequency noise, repetitive textures) and significantly reducing misclassifications in this problem domain. Using a five-fold cross-validation and a dataset of 2,003 images in six classes, an average balanced accuracy of 99.26 % can be achieved, outperforming the baseline model (98.50 %) and thereby reducing misclassifications by approximately 46 %. We demonstrate that rather than increasing model complexity, leveraging domain-specific characteristics can lead to highest accuracies. This will enable the application of automatic railway surveillance in the future to ultimately enhance security.
Funding Information
  • Open-Access-Publication-Fund of the Helmut-Schmidt-University/University of the Federal Armed Forces Hamburg

This publication has 61 references indexed in Scilit: