In the recent progress towards autonomous trains, the Automated Driving System (ADS) provides the so-called perception units. These components implement Artificial Intelligence (AI) modules for the assessment of events and their relevance around the system to ensure the situational awareness. This raises many challenges concerning the level of safety (globally at least equivalent to the level of the existing system with a human operator) and the means to achieve it for safety-critical transportation. Our work aims to clear the pathway of AI systems performances meeting the safety-related requirements of autonomous train. In this paper, we propose a methodology of elicitation of safety requirements from system level to AI model level taking into consideration a use case related to the perception function: monitoring of Level Crossing (LX) failure.