Safety-Critical Scenarios for Autonomous Driving: A Survey of Methods, Benchmarks, and Verification Pipelines

Abstract

The safe deployment of autonomous vehicles depends on their ability to perceive, handle, and validate safety-critical scenarios, namely rare but complex situations that are highly relevant to real-world safety risks. Different from previous surveys on safety-critical scenarios in autonomous driving, this survey systematically reviews safety-critical scenarios from three perspectives: methods, benchmarks, and verification pipelines. We first introduce a five-domain taxonomy that organizes existing methods from geometry-based safety estimation and deep risk anticipation to risk grounding, uncertainty-aware control, and integrated control with world models and vision-language models. We then review traffic accident datasets and critical scenario benchmarks. Based on these datasets and benchmarks, we further derive thirty representative critical scenarios for evaluation and benchmark design. Finally, we summarize verification pipelines and outline future directions. The review shows that current studies are moving toward closed-loop safety assurance, but standardized benchmarks, uncertainty-aware control, and traceable validation evidence remain insufficient.