Genetic Algorithm-Based Challenging Scenarios Generation for Autonomous Vehicle Testing

Virtual testing plays a vital role in developing connected and automated vehicles. To ensure the safety of CAV, millions of miles of testing are required. The scenario library defines the parameter of each testing scenario, which reduces testing by merging similar scenarios and focusing on challenging scenarios. However, current scenario generation methods still bear some disadvantages, such as lack of data and low efficiency. In order to overcome the lack of data, this paper proposed a genetic algorithm-based framework for the cut-in event and car-following event. In this study, the genetic algorithm searches for one of the most challenging scenarios within the boundaries. The challenging scenario is acquired by the brute force search algorithm which searches from the start scenario. The result shows the ratio of accident scenarios to challenging scenarios in our scenario set is more than 40%. Compared with natural driving data scenarios, our method performs well without natural driving data. In addition, the genetic algorithm converges in ten iterations which indicates its efficiency and suitability in scenario generation. More specifically, whole challenging scenarios which probably occur in the real world are collected in the scenario set.