Enhancing Rail Obstacle Detection Systems: Optimizing Accuracy in Adverse Weather Conditions

Abstract

Railway safety is paramount, especially with the increasing reliance on rail transport and the potential for catastrophic consequences from train colliding with obstacles. This paper introduces a novel obstacle detection methodology using Convolutional Neural Networks (CNNs) to enhance detection accuracy, particularly for diverse and unforeseen obstacles, including wildlife intrusion, under challenging environmental conditions. We employ the state-of-the-art (You Only Look Once) YOLOv11-Seg algorithm for simultaneous rail segmentation and obstacle detection, defining a critical safety margin around the tracks. A key contribution of this work is a novel synthetic image generation algorithm designed to address the critical scarcity of real-world obstacle data, particularly for rare and unpredictable hazards such as animals and uncharacterized debris. This algorithm strategically places various obstacles, extracted from diverse sources, at random locations on the rail or within the safety margin. Crucially, it incorporates diverse and realistic environmental conditions, such as train vibrations, rain, snow, dust, fog, and varying light intensities to augment the training data and improve the model’s robustness against these highly transient events. Experimental results demonstrate the effectiveness of the YOLOv11-Seg network, trained on our synthetically augmented data set, in accurately performing both segmentation and obstacle detection in a single step, paving the way for improved railway safety systems.