HITP: Hierarchical Intention Representation for Multi-Agent Trajectory Prediction

Multi-agent trajectory prediction is a fundamental task in autonomous driving, requiring accurate understanding of the latent intentions behind road user motions. However, capturing both long-term goals and short-term dynamic preferences in complex, multi-agent environments remains highly challenging. In this paper, we introduce a novel Hierarchical Intention-based Trajectory Prediction (HITP) framework that explicitly models agent intentions across different temporal scales through a frequency-aware, multi-modal representation. By decomposing historical trajectories into the frequency domain, HITP disentangles long-horizon navigational semantics from short-term interaction dynamics. Structured intention features are then extracted via attention-based encoders tailored for goal planning and motion preference modeling. To integrate these hierarchical intentions with road topology and inter-agent relations, we design a relational fusion decoder guided by structured relational encodings. To further enhance intention learning, we introduce two auxiliary tasks: long-term goal regression to supervise destination planning and masked trajectory modeling to regularize local motion understanding. Extensive experiments on the Argoverse motion forecasting benchmark demonstrate that HITP achieves competitive prediction accuracy while offering improved interpretability and social compliance. Overall, HITP provides a unified framework for structured, interpretable, and map-consistent multi-agent trajectory forecasting in complex urban scenarios.