Quantile-based scenario generation for automated vehicle safety evaluation

As automated vehicles (AVs) are increasingly deployed, ensuring their safety and reliability is crucial before widespread adoption. Existing safety evaluation methods typically focus on generating a testing scenario library with a large number of safety-critical scenarios; however, this approach presents two key limitations. First, the safety testing of the testing scenario library is time-consuming, making it impractical to apply to the production qualification test for every individual production AV. Second, most methods aim to maximize the risks of the scenario, often overlooking that some highly hazardous situations are unavoidable. Considering these research gaps, this study introduces a quantile-based scenario generation method for AV safety evaluation. The proposed method generates scenarios with varying levels of risk, determined by a specified quantile of the risk index, enabling a comprehensive and efficient assessment of AV safety. With the knowledge of the quantile of the scenario library, safety evaluation can rapidly identify the safety performance of each individual AV with a theoretical bound using a limited number of tests. To address the challenge posed by the rarity of safety-critical events, an adaptive variance reduction framework based on importance sampling theory, combined with Particle Swarm Optimization, is employed to minimize estimation variance and optimize scenario distribution. Experiments validate the method's ability to reduce estimation variance in the multi-lane scenario and demonstrate how it compares the safety performance of commercialized AVs.