Adaptive graph transformer with future interaction modeling for multi-agent trajectory prediction

Abstract

Forecasting the trajectories of traffic agents is essential for autonomous systems such as self-driving cars and social robots to guarantee safety in crowded scenarios. Capturing social interactions between agents and generating informative future features bring great challenges to accurate trajectory prediction. To this end, this paper proposes a novel multi-agent trajectory prediction model called AGTFI based on the adaptive graph transformer and future interaction modeling. First, an adaptive graph transformer (AGT) proficient at extracting node and edge features is introduced to capture the complex social interactions between traffic agents. Moreover, a two-stage prediction approach is devised where the first stage is devoted to generating pre-estimated future motion features by bidirectional corrected GRU (BCGRU) and the second stage further incorporates future social interactions into BCGRU to reduce prediction errors. Quantitative and qualitative evaluations of AGTFI on benchmark datasets, including ETH-UCY, SDD, and INTERACTION demonstrate the effectiveness of our model. Ablation studies are conducted to verify the rationale behind the model components.