Jintao, Huang (Southeast university), Guo, Yanyong (Southeast University), Yuguang, Chen (Southeast university), Wu, Hao (Southeast University)

ST-DiffNav: Spatio-Temporal Conditional Diffusion with Navigational Priors for Autonomous Trajectory Prediction

Scheduled for presentation during the Regular Session "S31a-AI-Driven Motion Prediction and Safe Control for Autonomous Systems" (FR-LM-T31), Friday, November 21, 2025, 11:10−11:30, Southport 1

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

To accurately model the complex interactions among dynamic traffic participants and adapt to the stochastic motion characteristics of heterogeneous agents, this paper proposes ST-DiffNav, a conditional diffusion model-based, environment-adaptive trajectory prediction framework for forecasting the future trajectories of all agents in a scene. The framework leverages a multimodal fusion strategy to effectively integrate historical motion trajectories, high-definition map semantics, and navigational priors, thereby enabling joint modeling of the spatiotemporal dynamics and learning structured trajectory distributions. To enhance training and inference efficiency, we introduce mixed-precision training to accelerate model convergence and adopt DPM-Solver to reduce the number of denoising steps, significantly improving inference speed without compromising prediction accuracy. This approach effectively mitigates the efficiency bottleneck commonly associated with diffusion models during both training and inference stages. We conduct extensive evaluations on the nuScenes dataset, and experimental results demonstrate that ST-DiffNav outperforms various state-of-the-art baseline models in trajectory prediction accuracy, validating its deployability in computationally constrained environments and its strong potential for real-world applications.