Diffutory: A Future Feature and Mode Association Augmented Diffusion Model for Trajectory Prediction in Autonomous Vehicles

Accurate prediction of future trajectories for surrounding agents is crucial for ensuring the safety and reliability of autonomous driving systems. Recent advances have witnessed the widespread adoption of generative models, particularly diffusion models, to enhance trajectory prediction. However, there still exist several challenges, particularly in terms of inadequate conditional guidance and insufficient mode association. In this paper, we propose Diffutory, a future feature and mode association augmented diffusion model for trajectory prediction. Our model operates in two principal stages, namely the distribution learning stage and the indeterminacy denoising stage. During the first phase, Diffutory carefully choreographs ample conditional guidance by modeling historical and forthcoming features of traffic scenarios. The important future feature encoder enables the model to envision potential future driving scenarios, facilitating a deeper understanding of agent dynamics and effectively improving predictive accuracy. In the second stage, a conditional denoising model is leveraged to progressively reduce noise and generate more accurate future trajectories. Furthermore, to correlate predicted modes while promoting diversity in generated trajectories, we introduce a novel mode association constraint. It augments the predictive capacity of our diffusion model without adding any additional computation cost. Extensive experiments on two real-world datasets provide compelling evidence of the superiority and effectiveness of our approach, demonstrating its ability to produce accurate and reliable trajectory predictions in dynamic and complex traffic environments.