SceGAN: A Method for Generating Autonomous Vehicle Cut-In Scenarios on Highways Based on Deep Learning

Abstract

With the increasing level of automation of autonomous vehicles, it is important to conduct comprehensive and extensive testing before releasing autonomous vehicles into the market. Traditional public road and closed-field testing failed to meet the requirements of high testing efficiency and scenario coverage. Therefore, scenario-based autonomous vehicle simulation testing has emerged. Many scenarios form the basis of simulation testing. Generating additional scenarios from an existing scenario library is a significant problem. Taking the scenarios of a proceeding vehicle cutting into an adjacent lane on highways as an example, based on an autoencoder and a generative adversarial network (GAN), a method that combines Transformer to capture the features of a long-time series, called SceGAN, is proposed to model and generate scenarios of autonomous vehicles on highways. An evaluation system is established to analyze the reliability of SceGAN using discriminative and predictive scores and further evaluate the effect of scenario generation in terms of similarity and coverage. Experiments showed that compared with TimeGAN and AEGAN, SceGAN is superior in data fidelity and availability, and their similarity increased by 27.22% and 21.39%, respectively. The coverage increased from 79.84% to 93.98% as generated scenarios increased from 2,547 to 50,000, indicating that the proposed method has a strong generalization capability for generating multiple trajectories, providing a basis for generating test scenarios and promoting autonomous vehicle testing.