Information Theory Based Quantitative Complexity Analysis for Autonomous Vehicle Safety Testing

Ensuring the safety and reliability of autonomous vehicles (AVs) requires rigorous scenario-based testing to evaluate their performance in dynamic and unpredictable environments. The ability to simulate diverse traffic conditions enables researchers to identify vulnerabilities and enhance system robustness prior to real-world deployment. Traditional AV safety testing relies on randomized or manually selected test cases, which may not prioritize high-risk scenarios that are more likely to reveal safety-critical issues. To optimize testing efforts, a systematic approach to scenario selection and prioritization is necessary.In this paper, we present an entropy-based methodology to quantify the complexity of driving scenarios, offering a structured and data-driven approach to AV safety testing. By computing quantitative complexity scores, our method ranks test scenarios based on their inherent difficulty, providing a scalable mechanism to improve test coverage. Compared to conventional scenario selection approaches, our methodology introduces a modular and adaptable framework that seamlessly integrates into existing AV validation pipelines. The approach supports multiple actor trajectories and is released as an open-source module, facilitating broader adoption. To demonstrate its effectiveness, we apply the complexity estimation framework to a realistic multi-agent scenario involving cars, a bicycle, and a pedestrian, analyzing scenario complexities over time. The entropy-based complexity score serves as a relative metric for assessing scenario difficulty, guiding structured AV testing and validation. By improving the precision and efficiency of scenario selection in AV safety testing, the approach contributes to the development of more reliable and resilient autonomous systems.