Lane-Aware Transformers for Multi-Agent Trajectory Prediction

Abstract

For autonomous driving vehicles, accurately predicting the future trajectories of interactive road agents and planning a trajectory that complies with societal requirements and resembles human-like behavior is extremely important. Existing multi-vehicle trajectory prediction methods have redundancy when dealing with multi-agent scenarios, that is, they repeatedly encode invariant scenes around each vehicle, such as lane lines, which leads to increased delays in the model's reasoning. To solve this problem, we propose a novel multi-agent trajectory prediction model based on space-time Transformer. The model models a driving scene as a heterogeneous graph with nodes representing traffic participants or road elements and edges representing semantic relationships between them. In terms of spatial relation coding, the coordinate information of nodes and their edges is no longer set in a fixed global reference system, but transformed into a node-centered local coordinate system. The final model can predict both the target lane segment and the corresponding future trajectory of the agent. Our extensive experiments on the Argoverse real dataset confirm that our algorithm not only works, but also has high accuracy.