Improving generative trajectory prediction via collision-free modeling and goal scene reconstruction

Abstract

In the context of bird’s-eye view traffic scenarios, accurate prediction of future trajectories for various traffic agents (e.g., pedestrians, vehicles) is crucial for driving safety and decision planning. In this paper, we present a generative trajectory prediction framework that incorporates both collision-free modeling and additional reconstruction for future scene context. For the social encoder, we leverage the collision prior by incorporating collision-free constraints (CFC). We construct a social model composed of multiple graphs, where each graph points to the collision prior calculated from uniform direction sampling. In the scene encoder, we employ an attention module to establish connections between trajectory motion and all scene image pixels. Additionally, we reconstruct a goal response map (GRM) aligned with the intended one, thereby enhancing the scene representations. Experiments conducted on nuScenes and ETH/UCY datasets demonstrate the superiority of the proposed framework, achieving a 13.8% reduction in off-road rate on nuScenes and an average 13.2% reduction in collision rate on ETH/UCY datasets.