Toward Multimodal Vessel Trajectory Prediction by modeling the distribution of modes

Abstract

Vessel trajectory prediction using AIS data plays an important role in maritime navigation warning and safety. A key aspect of trajectory prediction is multimodal because of the uncertainty of vessel behavior. However, complex trajectory modes are difficult to be learned from low-dimensional AIS data with noise. In this paper, we propose a new method for multimodal vessel trajectory prediction, called Multimodal Vessel Trajectory Prediction via Modes Distribution Modeling (VT-MDM). This approach addresses the above challenges by introducing additional hiding regimes to characterize complex trajectory modes independently. Specifically, we introduce an additional latent vector as the encoding of the trajectory modes, which is randomly sampled from a multivariate Gaussian distribution to generate multiple predicted trajectories. To enable this Gaussian distribution for capturing the vessel trajectory modes, we use adversarial learning to enforce all its realizations to generate realistic predicted trajectories. Furthermore, we also encourage the mapping between the latent vectors of the modes and the predicted trajectories to be invertible and smooth, which prompts VT-MDM to produce truly and gradually multimodal predicted trajectories. Experiments on the real AIS dataset show that our method is capable of multimodal trajectory prediction with high accuracy.