Parameter Selection of Safety-Critical Scenarios Based on Scenario Database and Generative Models and Its Application to Generation of Car-Following and Cut-in Scenarios 시나리오 데이터베이스와 생성형 모델을 이용한 상세 시나리오의 파라미터 선정과 선행차량 추종 및 끼어들기 시나리오 생성으로의 적용

Abstract

The selection method of concrete scenarios based on scenario database and generative models is proposed in this paper. Two sampling methods are used in selecting appropriate parameters along a performance measure. First, the parameter space is extracted from scenario DB, and a set of parameters are selected via random sampling. Once the set of concrete scenarios are simulated, their distribution is analyzed with respect to a specific measure. The second method is based on parameter generative models. Simulated scenarios are used to train a surrogate model, Which is a multi-layer perceptron model. Then, a generative model is designed to search for the desired parameter based on the surrogate model. Thus, the second one can be used to condensate for the imbalance in randomly sampled concrete scenarios. Finally, the proposed selection method is more efficient than random sampling from the viewpoint of the distribution of two different measures.