Vehicular Multimodal Motion Forecasting With Scenario-Guided Diffusion Probabilistic Models

Ensuring the safe driving and decision-making of autonomous vehicles requires reliable modeling of the multimodal motion of vehicles. In this paper, we propose a novel scenario-guided diffusion probabilistic model called RoadDiff to address these issues. By modeling and interacting with scene-contextual heterogeneous data at the feature level, we achieve lightweight and high-quality representations of traffic dynamics. We interpolate these representations as guiding conditions into a stochastic iterative denoising process of reverse diffusion, progressively transforming a known prior distribution into a multimodal trajectory distribution that includes the ground truth mode. Our well-designed Transformer-based neural network approximates the denoising matching function of the reverse diffusion process, and our region of interest-based training strategy ensures that the sampled multimodal trajectories conform to road and physical constraints to overcome mode blur. We carefully design two loss functions to ensure multimodal prediction for regions of interest and efficient trajectory guidance during this process. Experimental results on real-world motion datasets demonstrate that RoadDiff outperforms state-of-the-art methods in terms of prediction multimodality and accuracy, indicating its potential as a novel paradigm for vehicular multimodal motion prediction.