Vehicle Trajectory Prediction

Autonomous driving is not exclusively a perception nor planning problem. A prediction pillar is essential and has been over-looked for years. While driving, humans have this powerful capability to anticipate other drivers’ decisions. Similarly, an autonomous vehicle should have the same prediction capability to safely navigate alongside human drivers. In this project, we study “generalizable” and “accurate” vehicle trajectory prediction solutions from different aspects:

  • Adding knowledge to the models: Adding human knowledge to the data-driven models is an effective approach to improve models’ generalization. To this end, we propose “Realistic Residual Block” (RRB) to combine any off-the-shelf knowledge-driven predictor with data-driven predictors. Link to the project