Abstract
Accurate prediction of human or vehicle trajectories with good diversity thatcaptures their stochastic nature is an essential task for many applications.However, many trajectory prediction models produce unreasonable trajectorysamples that focus on improving diversity or accuracy while neglecting otherkey requirements, such as collision avoidance with the surrounding environment.In this work, we propose TrajDiffuse, a planning-based trajectory predictionmethod using a novel guided conditional diffusion model. We form the trajectoryprediction problem as a denoising impaint task and design a map-based guidanceterm for the diffusion process. TrajDiffuse is able to generate trajectorypredictions that match or exceed the accuracy and diversity of the SOTA, whileadhering almost perfectly to environmental constraints. We demonstrate theutility of our model through experiments on the nuScenes and PFSD datasets andprovide an extensive benchmark analysis against the SOTA methods.