Boundary Scenario Generation for HAVs Based on Classification and Local Sampling

Abstract

High-level Automated Vehicles (HAVs) are expected to improve traffic safety significantly. However, verifying and evaluating HAVs remains an open problem. Scenario-based testing is a promising method for HAV testing. Boundary scenarios exist around the performance boundary between critical and non-critical scenarios. Testing HAVs in these boundary scenarios is crucial to investigate why collisions cannot be avoided due to small changes in scenario parameters. This study proposes a methodology to generate diverse boundary scenarios to test HAVs. First, an approach is proposed to obtain at least one High-Performance Classifier (HPC) based on two classification algorithms that iteratively guide each other to find uncertain scenarios to improve their performance. Then, the HPC is exploited to find candidate scenarios highly likely to be boundary scenarios. To increase the efficiency of candidate scenario generation, a strategy based on local sampling is presented to find more diverse candidate scenarios based on a small number of them. Numerical experiments show that the HPCs acquired by the method proposed in this study can achieve a classification accuracy of 98% and 99% for random car-following and cut-in scenarios, respectively. Moreover, more than 86% of 271,744 candidate cut-in scenarios derived by local sampling are near the performance boundary.