Driving-Over Detection in the Railway Environment

2026-02-19

AI Summary

This study addresses the lack of reliable automated detection methods for wheel-rail crushing events—such as wheels running over foreign objects on tracks—in fully automated train operations. It presents the first systematic investigation into automatic recognition of such events in railway scenarios. A dataset was constructed through controlled crushing experiments involving objects of multiple materials, and the authors propose and compare a convolutional neural network (CNN)-based detection approach against two classical thresholding methods. Experimental results demonstrate that the proposed CNN model achieves an average accuracy of 99.6%, substantially outperforming the threshold-based methods, which attain accuracies of 85.0% and 88.6%, respectively. These findings confirm the superiority and feasibility of deep learning for this critical railway safety task.

Technology Category

• Computer Vision: Motion & Tracking
• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Machine Learning: Learning on the Edge & Model Compression

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications

Abstract

To enable fully automated driving of trains, numerous new technological components must be introduced into the railway system. Tasks that are nowadays carried out by the operating stuff, need to be taken over by automatic systems. Therefore, equipment for automatic train operation and observing the environment is needed. Here, an important task is the detection of collisions, including both (1) collisions with the front of the train as well as (2) collisions with the wheel, corresponding to an driving-over event. Technologies for detecting the driving-over events are barely investigated nowadays. Therefore, detailed driving-over experiments were performed to gather knowledge for fully automated rail operations, using a variety of objects made from steel, wood, stone and bones. Based on the captured test data, three methods were developed to detect driving-over events automatically. The first method is based on convolutional neural networks and the other two methods are classical threshold-based approaches. The neural network based approach provides an mean accuracy of 99.6% while the classical approaches show 85% and 88.6%, respectively.

Problem

• driving-over detection
• railway environment
• automatic train operation
• collision detection

Innovation

• driving-over detection
• convolutional neural networks
• automatic train operation
• railway safety
• threshold-based methods