TIRP: a topology-informed refinement model for multimodal trajectory prediction in autonomous driving

Abstract In highly dynamic urban environments, accurate forecasting of surrounding vehicle trajectories remains critical for ensuring safety in autonomous driving vehicles. Despite recent advances, current methods struggle to capture the complex interplay between structured road networks and multi-agent interactions, particularly in challenging traffic scenarios. To tackle this challenge, we propose TIRP, a topology-informed refinement model for multimodal trajectory prediction in autonomous driving. This model addresses key limitations through three core components. First, the model ingests vehicle historical trajectories and high-definition maps into a prediction embedding module. This component extracts rich static road semantic information from vectorized maps using a three-channel skip graph convolutional network. Second, we design a hierarchical attention interaction network enhanced with dual-path hypergraph attention to model high-order interactions among traffic participants. Finally, the trajectory generation module employs a two-stage refinement decoder that integrates spatio-temporal context information and interaction features to produce multimodal trajectory outputs, progressing from coarse to fine-grained predictions. Extensive experiments on the Argoverse 1.1 dataset show that our method consistently outperforms prior state-of-the-art methods on the primary evaluation metrics; under the K = 6 setting, it achieves a minADE of 0.728, a minFDE of 1.154, and an MR of 0.103. The model effectively mitigates mode collapse while enhancing prediction reliability in heterogeneous road networks and dense traffic conditions. This advancement provides robust perceptual support for autonomous driving planning systems, offering substantial practical implications.