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

Zhaoxin Su, Gang Huang, Zhou Zhou, Yongfu Li, Sanyuan Zhang, Wei Hua

Abstract: Highlights•A generative trajectory prediction framework is presented.•Collision-free constraints are incorporated to guide social modeling.•An attention module establishes connections between motion and all scene pixels.•A response map combining goal information with relevant scene detail is used.•The proposed method is validated on nuScenes and ETH/UCY datasets.