Multi-modal Vehicle Trajectory Prediction via Attention-based Conditional Variational Autoencoder

Predicting the motion of surrounding vehicles is an essential ability for autonomous vehicles in real traffic, which can be beneficial to improving motion planning and decisionmaking. However, vehicle trajectory prediction has been a challenging task in autonomous driving, due to the uncertainty of driving intention and the interactions among multiple agents, etc. In this paper, we propose a novel encoder-decoder model via attention-based conditional variational autoencoder (CVAE), which is designed to predict multi-modal trajectories of vehicles on freeways according to different maneuvers. This model employs a multi-agent attention module for robustly capturing spatial interactions, and soft attention mechanism is used to learn temporal dependencies of trajectories. Besides, scheduled sampling is applied in the decoding process for improving the convergence of sequence prediction. We compare our approach with the prior methods on the NGSIM US-101 and I-80 datasets. The experimental results show that our model outperforms the existing methods in terms of the root mean square error (RMSE). Extensive experiments are performed to demonstrate the effectiveness of our proposed model.