TPNet: Trajectory Proposal Network for Motion Prediction
Liangji Fang1*  Qinhong Jiang1*  Jianping Shi1  Bolei Zhou2 
1SenseTime Group Limited
2The Chinese University of Hong Kong
*Equal contribution.
The movements of traffic agents are often regularized by the movable areas, while there might be multiple plausible future paths for the agents. Based on these prior knowledge, we propose a novel two-stage motion prediction framework, called TPNet, both for vehicles and pedestrians. By steering the proposal generation process, safe and multimodal predictions are realized. More importantly, the two-stage pipeline is flexible to encode different prior knowledge into the deep learning method.
We show the effectiveness of two-stage pipeline, diversity, safety.
Check more results in the following video.
  title     = {Trajectory Proposal Network for Safe and Multimodal Motion Prediction},
  author    = {Fang, Liangji and Jiang, Qinhong and Shi, Jianping and Zhou, Bolei},
  booktitle = {CVPR},
  year      = {2020}