A Multi-Dimensional Coverage Metric with Evolutionary Search for Safety-Critical Scenario Generation in Autonomous Driving Testing

Ensuring autonomous driving safety requires rigorous testing across diverse and safety-critical scenarios. Manual scenario design is labor-intensive and insufficient in capturing edge cases, while random generation produces redundant test cases. This paper proposes a coverage-guided evolutionary search algorithm (CGES) for automated generation of safety-critical test scenarios with quantitative coverage assessment. A parameterized scenario representation is established based on six functional dimensions, and three complementary coverage metrics---scenario parameter space coverage (SPSC), behavioral diversity coverage (BDC), and risk-weighted fault mode coverage (RFMC)---are defined to quantify test adequacy. An adaptive evolutionary search strategy that incorporates risk-prioritized fitness evaluation and diversity-aware selection is designed to efficiently explore high-risk regions of the scenario space. Experiments on CARLA using five operational design domains demonstrate that CGES achieves 17.3% higher composite coverage and discovers 28.6% more unique safety violations than the baselines, while reducing redundant test cases by 41.2%. The proposed metrics provide a quantitative foundation for evaluating the completeness of autonomous driving test suites, contributing to standardized safety validation aligned with NHTSA regulatory requirements.