Context-Based Vessel Trajectory Forecasting: A Probabilistic Approach Combining Dynamic Bayesian Networks with an Auxiliary Position Determination Process

This paper introduces a probabilistic approach for forecasting vessel trajectories. It combines a Dynamic Bayesian Network (DBN) and an auxiliary position determination process to iteratively sample future vessel positions in a scalable and computationally efficient manner. The DBN is a discrete probabilistic model of typical vessel behaviors. It is used for ancestral sampling to predict the speed and orientation of a vessel which, in turn, are used by the auxiliary process to predict the vessel's position in a discretized representation of the space. The DBN is event based and uses latent variables that efficiently encode the context influencing the dynamics of different types of vessels. The parameters of the DBN are learned in an unsupervised fashion by using the Expectation Maximization (EM) algorithm. The experiments with real world data confirm the accuracy and effectiveness of the proposed approach.