Learning Representative Vessel Trajectories Using Behavioral Cloning

Abstract

We suggest a data-driven approach to predict vessel trajectories by mimicking the underlying policy of human captains. Decisions made by those experts are recorded by the automatic identification system (AIS) signals and can be fused with additional non-kinematic factors like destination, weather condition, current tide level or ship size to get a more accurate snapshot of the situation that led to chosen maneuvers. In this work, we explore the usage of a method meant for optimal control, namely Behavioral Cloning, in a forecasting problem, in order to generate end-to-end vessel trajectories purely based on a given initial state. The training and test datasets consist of trajectories from the coast of Bremerhaven, having more than one thousand unique ships and different motion clusters. These are processed by a single deep-learning model, showing promising results in terms of accuracy and providing a research avenue for a near real-time application where vessel trajectories are to be forecast from a given snapshot of the situation - not from the costly history of all the vessels present.