Abstract:

To enhance the overall realism of background interactive vehicle trajectories in digital simulation scenarios for autonomous driving, this study approached the problem from both microscopic and macroscopic perspectives. Firstly, vehicle trajectory prediction models were trained on naturalistic driving data. Leveraging the characteristic that model-predicted trajectories more closely resembled real-world vehicle trajectories, the prediction served as the artificial intelligence (AI) driver model for background vehicles in simulation environments, improving the microscopic realism of simulated vehicle trajectory interactions. Building on this foundation, a measurement method for trajectory feature parameter statistical distribution differences and a corresponding optimization algorithm were designed, to re-select a single trajectory with the highest probability from multiple multi-modal prediction outputs, as the final driving trajectory for simulated vehicles, further enhancing the macroscopic realism of the generated trajectory feature parameter statistical distribution. The results show that, based on the proposed measurement metrics, the distribution difference between optimized simulated trajectories and real trajectories is reduced by 56.29% compared to pre-optimization, effectively enhancing the realism of background vehicle trajectories in simulation scenarios.

Key words: multimodal trajectory prediction, trajectory snapshot, trajectory feature vector clustering, Kullback-Leibler (KL) divergence, Bayesian optimization