Vehicle Trajectory Prediction Using Intention-based Conditional Variational Autoencoder

Xidong Feng, Zhepeng Cen, Jianming Hu, Yi Zhang

Abstract: Vehicle trajectory prediction has been an active research area in autonomous driving. In a real traffic scene, autonomous vehicle needs to predict future motion of surrounding vehicles before motion planning to improve driving safety and efficiency. In this paper, we first modify a sequence to sequence (seq-to-seq) maneuver-based model to produce possibility prediction of vehicle trajectory in future 5 seconds. Then we propose a novel method based on conditional variational autoencoder (CVAE). Our model generates multi-modal trajectory possibility prediction with high interpretability according to the estimation of driver's latent intention. Finally, we experiment the model on public traffic dataset and compare it with prior methods on trajectory prediction. The results show a great improvement on both lateral and longitudinal motion prediction, which also demonstrates the effectiveness of our model.