Risk-Aware Railway Obstacle Detection via Track Segmentation and Lateral Distance Estimation

Abstract

Existing railway obstacle detection methods primarily focus on object localization and classification, while lacking the capability to directly support operational risk assessment. To address this limitation, this paper proposes a risk-aware railway obstacle detection framework that integrates track segmentation and lateral distance estimation. The proposed framework jointly optimizes obstacle detection and track segmentation through shared feature learning, and estimates the lateral distance between detected obstacles and track boundaries based on the extracted track geometry. Railway clearance constraints are further incorporated for obstacle risk-level determination, thereby establishing an end-to-end pipeline from obstacle detection to operational risk assessment. Consequently, an end-to-end pipeline is established to classify obstacle risk levels in railway environments. To validate the proposed framework, a dedicated RN-rail-Object dataset is constructed, and comprehensive ablation studies, comparative experiments, and edge deployment evaluations are conducted. Experimental results demonstrate that the proposed method achieves 89.2% mAP for obstacle detection, 93.7% mIoU and 97.1% Dice for track segmentation, while maintaining an inference speed of 214.6 FPS on the RTX A4000 platform. Furthermore, edge deployment demonstrates the feasibility of integrating obstacle detection, track extraction, and geometry-based risk assessment on an edge computing platform, with representative scenarios illustrating the identification of different obstacle risk states. The proposed framework provides an interpretable geometry-based approach for extending conventional obstacle detection toward railway risk-aware perception, showing potential for intelligent railway inspection, early warning, and safety-oriented operation and maintenance.