ABSTRACT Autonomous driving safety validation increasingly relies on scenario‐based simulation. A feature‐knowledge and Wasserstein generative adversarial network (WGAN) co‐driven scenario generation method is proposed to address the issues of insufficient authenticity and limited scenario parameter dimensions in multi‐vehicle interaction scenario modelling. Key state parameters of cut‐in trajectories are extracted from natural driving data, combined with relative position and velocity information of background vehicles to form input vectors, thereby establishing a feature‐knowledge driven cut‐in scenario generation framework. To comprehensively evaluate the superiority of the proposed method, we conducted comparative experiments with five different generative adversarial network models as well as a diffusion model. Experimental results demonstrate that our proposed WGAN model reduces the JS divergence effectively, significantly outperforming other comparison models. Visualisation analysis further validates the distribution consistency and diversity of the generated scenarios. This study provides an effective solution for high‐dimensional scenario modelling and generation in autonomous driving simulation testing. Codes are open‐sourced at: https://github.com/MZJSJMC/MVIC‐Scenario‐WGAN# .
Mo et al. (Thu,) studied this question.