Residual Diffusion Model for Joint Stochastic Trajectory Prediction in Roadside Surveillance Environments

Roadside surveillance video signals usually contain historical trajectory information and environmental data of vehicles in a specific environment. Vehicle trajectories can be defined as time series signals to facilitate the prediction of future vehicle trajectories. We propose a method based on a residual diffusion model to reason about the joint distribution of future trajectories across multiple vehicles. This approach has several key advantages: First, the model can learn multiple probability distributions that capture different potential future outcomes for multiple vehicles. Secondly, by combining the trajectory information of multiple vehicles, the model can reason in the way of the standard denoising model and multiple residual denoising models, so as to improve the model performance and prediction speed. Finally, a general constraint function was introduced to ensure the control trajectory of multiple vehicles and avoid collisions. A large number of experimental results on the NGSIM dataset show that the model has a significant improvement in prediction accuracy compared with the baseline method.