Attention Enhanced Transformer for Multi-agent Trajectory Prediction

Precise trajectory prediction constitutes a pivotal challenge in the field of autonomous driving, which motivates the integration of interactive dynamics among surrounding agents and the geographical map information for better trajectory predictions. However, the existing Graph Neural Network (GNN) based methods constructed the graphs based on Euclidean distance to determine the existence of interactions, which might introduce noise and trivial interactions leading to decreased performance and increased computational cost. Furthermore, most existing methods employ Gaussian or Laplace Mixture Models to cater to multimodality distributions, without considering how to generate multimodal features more effectively. To address these challenges, we propose the Attention Enhanced Transformer for dealing with multi-agent trajectory prediction task (AET). Specifically, by introducing a causal attention mechanism, we partition the original graph into a causal attended graph and a trivial attended graph, thereby enhancing inference speed and accuracy. Additionally, by introducing a multimodal attention mechanism, we are able to allocate attention scores more reasonably, thereby obtaining more distinct multimodal trajectories and more accurate trajectories for each modality. Experiments demonstrate that AET achieves State-of-the-Art performance on INTERACTION and Argoverse datasets.