Authors:
Mohamed Elgharbawy, Andreas Schwarzhaupt, Michael Frey, Frank Gauterin
Keywords:
field operational tests, operational design domain, sensitivity and reliability analysis, Monte-Carlo and adaptive importance sampling methods
Abstract:
This paper presents a data-driven risk-based framework to predict the safety confidence of automated driving functions using sensitivity and reliability analyses. Due to the rare nature of accidents, the Field Operational Tests will require prohibitive number of driving hours to prove the statistical superiority. Owing to the complexity of the automated driving functions, an oracle for safety certification is unlikely to be provided by a single test method. For this reason, our approach envisages the identification of failure types to breakdown the functional complexity. On one hand, a systematic process to complement virtual testing is demonstrated by extracting knowledge from real-world testing database using time-series analysis with hierarchical clustering and normalized cross-correlation. On the other hand, the extracted clusters and their parameter space define the probability of occurrence of each logical scenario and the probability distributions of the associated parameters. The minimum time-to-collision is used as a prototypical safety measure. Thereby, the sensitivity analysis employs meta-model techniques, where the parameters are considered as input and the safety measure as output. Finally, the reliability analysis identifies the failure region in the parameter space to predict the probability of failure for each logical scenario using Monte-Carlo and Adaptive Importance sampling methods.
This article is available for DSC 2020 attendees. Please enter your ConfTool credentials:
A Driving Simulation Association membership is required in order to download the papers.