Train Operation Risk Identification Paradigm Based on Forward Vision Sensor

With the advancement of railway transportation toward higher speeds and greater density, environment perception technology has become crucial for ensuring operational safety. This paper proposes a novel risk identification paradigm for train operation based on forward-facing visual sensors, enabling real-time hazard awareness in the track area ahead through a multi-task deep learning network. The proposed method first introduces a parallel perception network architecture based on an improved YOLOv8 framework, which simultaneously performs obstacle detection and track semantic segmentation within a single network, achieving multi-dimensional environmental comprehension. Subsequently, a multi-feature risk space is constructed by integrating obstacle distribution, track coverage rate, and category-specific hazard coefficients, with a dynamic weight adjustment mechanism to adapt to diverse operational scenarios. Furthermore, a temporal risk evolution model is incorporated, enhancing the stability and reliability of assessment results through the joint analysis of instantaneous risk, smoothed risk, and risk trends. Case studies demonstrate that the system can rapidly issue warnings in pedestrian intrusion scenarios, while the comprehensive risk indicator efficiently and accurately reflects the dynamic evolution of threat severity.