Scenarios play an important role in virtual test of autonomous vehicles. In order to reduce the negative impact of design defects, the robustness of the software of autonomous vehicles must be adequately tested, which requires vast virtual scenarios having a high coverage on the realistic ones. Vehicle lane-change is a frequent traffic scenario with high risk on highways. As the contribution of this paper, a couple of features that affect the risk degree of lane-change scenario, including location and motion information, kinematic and geometric parameters of the ego and adjacent vehicles, were extracted from the open highD dataset, which is an high-definition trajectory dataset on German highways, to reconstruct the lane-change traffic scenario. Furthermore, the risk degree of lane-change scenario was modeled using TTC(Time to collision) as the quantitative criterion. At last, a mapping between data features and risk degree of lane-change scenario was established with Generative Adversarial Network (GAN), aiming to reconstruct the scenarios with the configurable risk degree. Simulation of scenario reconstruction shows that the rebuilt lane-change scenarios are capable to cover all scenarios in the open dataset. The proposed scenario reconstruction method can provide helpful supports to the virtual testing of autonomous vehicles. keywords: Autonomous Driving, Lane-change Scenarios, Trajectory, GAN

CICTP