Autonomous driving systems require extensive high-quality data for training and validation to ensure safety and reliability. However, acquiring such data is often inefficient, costly, and sometimes infeasible, particularly in complex interactive scenarios, necessitating effective trajectory data generation methods. Knowledge-based methods struggle to explicitly model the intricacies of vehicle interactions, such as competition and right-of-way negotiation. In contrast, data-driven models, while effective at capturing dynamic interactions, often fail to derive common-sense knowledge from real-world data, leading to unrealistic outcomes. To address these limitations, this study proposes the knowledge-infused variational autoencoder (KI-VAE) model for generating physically feasible and diverse vehicle trajectories. KI-VAE employs a GRU-augmented encoder for temporal features of vehicle trajectories. Linear interpolation merges safe and collision-free trajectory features, while style transfer ensures alignment with road constraints. The decoder, incorporating vehicle kinematic constraints, guarantees adherence to physical limitations. We evaluate the proposed model through direct metrics and an indirect evaluation task based on trajectory prediction performance. KI-VAE outperforms baselines including MTG and InfoGAN, achieving up to 5.3% improvement in smoothness (MDN) and ensuring compliance with acceleration constraints. Furthermore, trajectory prediction models trained on KI-VAE-generated data achieve up to 13.5% improvement in FDE and 6.8% in ADE, indicating superior diversity and generalizability. These results demonstrate KI-VAE’s effectiveness in generating physically plausible and prediction-enhancing trajectories, making it a promising solution for autonomous driving training and evaluation pipelines.
Chu et al. (Tue,) studied this question.