Zero-shot trajectory generation with self-supervised diffusion model

Abstract

Vehicle trajectory data play a crucial role in modern intelligent transportation systems. Traditional generative models, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), often struggle to accurately capture the distributional characteristics of trajectories in real-world scenarios. Moreover, conventional traffic trajectory generation methods face challenges due to their reliance on real-world road network data that are difficult to acquire and have limited availability. To address these issues, this study leverages the inherent correlation between surveillance camera data and vehicle trajectories, developing a zero-shot trajectory generation method (defined as generating valid trajectories without any complete single-vehicle trajectory annotations, only using aggregated traffic flow statistics and road network topology) based on a diffusion model. We treat surveillance traffic flow distributions as conditioning information to guide the denoising process continuously, employing an adaptive optimization framework to generate compliant and high-quality trajectories. All experiments are validated on real-world traffic datasets from Dadukou District, Chongqing. Notably, the entire model training process does not use any complete single-vehicle trajectory annotations, relying only on aggregated traffic flow statistics and public road network topology. Using surveillance traffic flow data as input, the generated synthetic trajectories not only align closely with real trajectories but also demonstrate comparable effectiveness. Extensive experiments validate the efficacy of the proposed method in zero-shot trajectory generation tasks, showing excellent performance in both trajectory similarity and distributional similarity.