As the railway domain progresses towards autonomy, maintaining safety at levels comparable to human-operated systems is a crucial challenge. Autonomous trains require advanced systems capable of real-time risk assessment and decision-making, a task traditionally managed by human situational awareness. This paper introduces a novel risk-based decision-making approach for autonomous trains, using Partially Observable Markov Decision Processes (POMDPs) for continuous monitoring and evaluation of environmental collision risks. By consistently maintaining an acceptable risk level through ongoing risk estimation (in terms of occurrence probability and severity degree), the approach supports the decision-making capabilities of the autonomous driving system in autonomous trains, enabling safe and informed decisions despite the uncertainties in the train’s operational state and environmental conditions. The approach’s relevance and effectiveness are illustrated through its application in an anti-collision function for autonomous trains.