Trajectory Prediction for Marine Vessels using Historical AIS Heatmaps and Long Short-Term Memory Networks*

Estimating the trajectory of other vessels is essential when navigating a marine vessel, both as a human navigator and as a machine. By estimating the trajectories of other vessels, sub-systems such as collision avoidance algorithms can plan ahead accordingly in order to avoid conflicts. To estimate the trajectories of other vessels, the use of Automatic Identification System (AIS) is a good candidate data-point, as this is becoming increasingly more common, and in some cases even mandated, on-board vessels. This paper presents a data-driven approach that uses the historical AIS data within a selected area in the Danish waters. The historical data is transformed into a probabilistic heat map using Kernel Density Estimation (KDE), and is further encoded using a Convolutional Autoencoder (CAE) before entered into the estimation scheme. The estimation scheme consists of a Long Short-term Memory (LSTM) model, in a Generative Adversarial Network (GAN) configuration, which is sampled multiple times, yielding a single trajectory prediction with uncertainty. The performance of the estimation scheme is demonstrated and compared against two other commonly used methods, showing that the probabilistic heat map provides valuable information, compared to the baseline methods.