EFIN-MP: Explicit Future Interaction Network for Motion Prediction

Accurate prediction the future movements of surrounding traffic participants is crucial for autonomous driving. Among various strategies, learning complex interactive behaviors among agents efficiently stands out as both an effective means and a challenging aspect of improving predictive model performance. To address this challenges, this paper proposes the Explicit Future Interaction Network for Motion Prediction (EFIN-MP), which aims for more precise agent motion predictions. By constructing a novel N-layer decoding module, EFIN-MP can effectively model future interactions between agents. Each layer capitalizes on the multimodal trajectory predictions from its preceding layer to explicitly integrate future interaction features. Specifically, In the first layer of the decoding module, EFIN-MP constructs a goal prediction module based on the attention mechanism, which regresses the multimodal predicted trajectories of multiple agents in the scene by fusing goal points with contextual features. To predict goal points efficiently, EFIN-MP establishes a dynamic search area in the preprocessing stage. This area integrates rules and knowledge, effectively utilizing road constraints and minimizing the number of potential goal points. Experiments show that EFIN-MP achieves outstanding performance on the Argoverse 1 motion forecasting benchmark and nuScenes dataset, demonstrating both high predictive accuracy and effectiveness.