Maritime anomaly detection using Gaussian Process active learning

A model of normal vessel behaviours is useful for detecting illegal, suspicious, or unsafe behaviour; such as vessel theft, drugs smuggling, people trafficking or poor sailing. This work presents a data-driven non-parametric Bayesian model, based on Gaussian Processes, to model normal shipping behaviour. This model is learned from Automatic Identification System (AIS) data and uses an Active Learning paradigm to select an informative subsample of the data to reduce the computational complexity of training. The resultant model allows a measure of normality to be calculated for each newly-observed transmission according to its velocity given its current latitude and longitude. Using this measure of normality, ships can be identified as potentially anomalous and prioritised for further investigation. The model performance is assessed by its ability to detect artificially generated AIS anomalies at locations around the United Kingdom. Finally, the model is demonstrated on case studies from artificial and real vessel data to detect anomalies in unusual tracks.