Research on the safety and security of autonomous trains for approval and operation is a multifaceted topic. It incorporates aspects as railway-specific legislation, technical norms, computer vision, machine learning, sensor technology, simulation environments, privacy protection, data quality standardization, human behavior and interaction prediction, and societal acceptance.
As in IEC 62267:2009, the levels of automation are on-sight train operation (GoA0), non-automated train operation (GoA1), semi-automatic train operation (GoA2), driverless train operation (GoA3), and unattended train operation (GoA4). Autonomous trains from GoA3 on (GoA3+) need automated obstacle detection. Current GoA3+ systems use human-run remote surveillance and control as a fallback system since modern computer vision is not yet reliable for that purpose. Even a reliable estimate of the failure rate is not secure. Current European railway-specific legislation requires a legally secure estimate of fatalities for the approval of autonomous trains and the approval precedes their regular operation. Such legally secure proof of safety for deep learning, the dominant technology for modern computer vision, requires a strong developer network, which elaborates generally accepted technical standards. Open datasets and open simulation environments promote such a developer network and therefore the enablement of autonomous trains. This talk summarizes present and future challenges on the way towards autonomous trains.
THIS KEYNOTE TALK REPRESENTS SOLELY THE AUTHOR’S PROFESSIONAL OPINION, NOT THE ONE OF HIS EMPLOYER