Incorporating safety relevance and realistic parameter combinations in test-case generation for automated driving safety assessment

Abstract

This research contextualizes within ongoing international efforts to harmonize a standardized methodology to evaluate automated driving systems' safety. Most methodologies currently under development generate test-cases by applying combinatory or stochastic approaches to scenario-specific vehicle kinematics parameter distributions extracted from naturalistic driving data. This implies that safety relevance and parameter correlations may not always be properly accounted for in generating the test-cases, and that the methodology itself becomes dependent on a reference safety model that needs to be agreed upon. Therefore, the specific aim of this paper is to develop a methodology that incorporates parameter correlations and accounts for safety relevance in the process to generate test-cases for automated driving systems safety evaluation, and that does not depend on reference models. First, we show the importance of incorporating parameter correlations within a given scenario, and propose a methodology to calculate linear dependencies in safety relevant situations. Second, we propose a methodology to generate test-cases that accounts for the previous correlation analysis results. In order to study the feasibility of the methodology proposed, we apply it to a set of cut-in maneuvers processed from naturalistic highway driving data previously collected in Germany. The developed methodology informs currently ongoing discussions on a possible way to approach test-case generation within the automated driving safety evaluation process.