Towards Realistic, Safety-Critical and Complete Test Case Catalogs for Safe Automated Driving in Urban Scenarios

The generation of parametrized test scenarios that prove the safety of automated driving system is still unresolved. The state of the art in data-driven test case generation presents more and more complex approaches, neglecting the goals of comprehensibility and practicality and hardly consider the requirement of realistic test case design. In this paper, we further develop our search-based test case generation methodology [1]. Among others, we introduce the novel approach of boundary functions where dependencies in the edge areas of a 2-dimensional parameter space are automatically detected and transformed into sampling restrictions for realistic test case design. Further, we apply the methodology on the unprotected left turn scenario based on an urban naturalistic driving dataset recorded in Germany [2]. Here, we present an advanced modeling approach and criticality metric that is suitable to generate test cases that evaluate the Vehicle under Test's capability to flexibly re-plan its driving trajectory during approaching. The generated test cases outperform a common sampling approach in terms of criticality and coverage and are applicable to comparable complex urban scenarios.