MFPINet: multi-feature fusion with progressive interaction for trajectory prediction in autonomous driving

Abstract Multimodal trajectory prediction is a crucial technology for the driving safety of autonomous vehicles. The core objective is to accurately infer the diverse motion intentions of traffic participants in future spatiotemporal domains, providing a reliable environmental perception foundation for decision-making. Current research primarily relies on single-stage interaction modeling paradigms based on graph neural networks and attention mechanisms. While these methods can capture local spatiotemporal correlations, they struggle to represent global interactions in complex scenarios fully. To address these challenges, this paper proposes a Multi-Feature Progressive Interaction Network for multimodal trajectory prediction. First, we construct a progressive interaction modeling network module, achieving deep interaction between traffic participants and road constraints in dynamic scenes through a staged feature interaction mechanism. In the primary interaction stage, a multi-head attention module is used to model the agent-agent local interaction relationships. In the secondary interaction stage, map encoding features are introduced to enable deep interaction between traffic participants and road constraints in dynamic scenes. Next, we design a decoder module with trajectory refinement capabilities, effectively integrating the interaction information between agents into the refined decoding features, enhancing the consistency between the predicted trajectories and the map. Finally, experiments are conducted on the authoritative Argoverse 1 dataset. Compared to existing state-of-the-art models, the proposed model achieves an average improvement of 16.5% and 6.9% on metrics minFDE and minADE , respectively. The results demonstrate that our model exhibits outstanding predictive performance in complex urban scenarios.