EURO 2021 Athens<br />Abstract Submission

Online presentation.

2569. Towards Risk Assessment of Learned Computer Vision for ATO

Invited abstract in session TC-50: AI4RAILS IV, stream The 2nd International Workshop on Artificial Intelligence for RAILwayS (AI4RAILS).

Area: Workshops and Satellite Events

Tuesday, 12:30-14:00
Room: Virtual Room 50

Authors (first author is the speaker)

1. Rustam Tagiew
Cross-cutting issues, digitalization, automation, migration and law, German Center for Rail Traffic Research
2. Thomas Buder
German Center for Railway Traffic Research
3. Kai Hofmann
German Center for Railway Traffic Research
4. Christian Klotz
German Center for Railway Traffic Research
5. Roman Tilly
German Center for Railway Traffic Research


Automatic Train Operation (ATO) from GoA3 on requires Computer Vision (CV). Machine Learning (ML) as the state of art for CV is still missing in binding technical norms for ATO and therefore requires an extra risk assessment policy for a certification. In this contribution, we consider a deterministic CV subsystem within conventional software and rule out end-to-end learning as well as learning during operation. This CV subsystem is not conventionally coded but learned on pre-recorded data solely ahead of the operation. Learned CV (LCV) neither guarantees zero systemic faults, nor can be fully covered in tests. Furthermore, LCV will be restricted to tighter system boundaries than with manual operation (MO) relying on human perception. The certification policy of the complete system incorporating LCV must at least ensure lower risk within system boundaries than of MO. Our project ATO-Sense will deliver quantitative results on the risk of MO and ATO-Risk quantitative results on how much lower the risk of ATO should be. Once a binding threshold for risk of ATO exists, diverse activities for risk assessment can be executed. This contribution includes an overview of such activities as cross-validation, Bayesian ML and statistics, PFMEA, MCDA, DoE, CV-HAZOP, reinforcement learning for testing, explainable AI and some other emerging techniques for deep learning models. THIS CONTRIBUTION REPRESENTS SOLELY AUTHORS’ PROFESSIONAL OPINION, NOT THE ONE OF THEIR EMPLOYER.


Status: accepted

Back to the list of papers