An Accelerated Safety Probability Estimation Method for Control Policies of Autonomous Vehicles in Open Environments

One of the critical bottlenecks that influence the development and deployment of autonomous vehicles (AV) is the safety assessment for raised control policies, especially in the situation of taking sensing noise into account. Safety probability estimation is one of the useful tools to describe the performance of control policies. However, these evaluation metrics always require amounts of tests to obtain, especially in open environments, which brings high time costs, no matter with actual vehicles or high-fidelity models. This article proposes an accelerated safety probability estimation method based on multi-fidelity models and a scenario classification method. Our method can accelerate the safety probability estimation process and has a small error. In our test of a desert road scene, our method used 1/750 of the time of the Airsim with Unreal Engine 4 (UE4), and get less than 5% error in most cases.