Conditional Wasserstein Auto-Encoder for Interactive Vehicle Trajectory Prediction

Trajectory prediction is a crucial task required for autonomous driving. The highly interactions and uncertainties in real-world traffic scenarios make it a challenge to generate trajectories that are accurate, reasonable and covering diverse modality as much as possible. This paper propose a conditional Wasserstein auto-encoder trajectory prediction model (TrajCWAE) that combines the representation learning and variational inference to generate predictions with multi-modal nature. TrajCWAE model leverages a context embedder to learn the intentions among vehicles and imposes Gaussian mixture model to reconstruct the prior and posterior distributions. Wasserstein generative adversarial framework is then used to match the aggregated posterior distribution with prior distribution. Furthermore, kinematic constraints are considered to make the prediction physically feasible and socially acceptable. Experiments on two scenarios demonstrate that the proposed model outperforms state-of-the-art methods, achieving better accuracy, diversity and coverage.