Domain Transfer of ODD Monitoring from Aviation to Automotive

Abstract

In autonomous systems, environmental conditions like fog can significantly degrade the reliability of vision-based artificial intelligence (AI) algorithms, often restricting their operation to a subspace of the operational domain called operational design domain (ODD). While fog serves as a critical example, this paper explores the broader concept of transferring ODD monitoring methodologies from aviation to the automotive domain. Through a case study, an existing image-based fog detection classifier, originally developed for unmanned aerial vehicles (UAVs), is adapted for application in the automotive sector. Using the CARLA simulation environment, the classifier is retrained and evaluated on a newly created, diverse dataset encompassing varying fog intensities and environments relevant to automotive scenarios. The results demonstrate the successful transfer of the fog detection classifier, suggesting the potential for cross-domain adaptation of algorithms that monitor operational domain attributes such as fog. Furthermore, this paper reviews and compares existing ODD monitoring approaches in both the aviation and automotive domains, highlighting their respective perspectives. This work contributes to understanding the similarities and differences in ODD monitoring across autonomous system domains.