Anomaly detection for sea surveillance

In this paper, unsupervised clustering of normal vessel traffic patterns is proposed and implemented, where patterns are represented as the momentary location, speed and course of tracked vessels. The learnt cluster models are used for anomaly detection in sea traffic. The Gaussian Mixture Model is used as cluster model and a greedy version of the Expectation-Maximization algorithm is used as clustering algorithm. The models have been trained and evaluated using real recorded sea traffic. A qualitative analysis reveals that the most distinguishing anomalies found in the traffic are vessels crossing sea lanes and vessels traveling close to and in the opposite direction of sea lanes. In order to detect complex anomalies involving multiple vessels and/or behavior that develop over time, a more sophisticated pattern model should be developed. Yet, the generality of the proposed model is stressed, as it is potentially applicable to other domains involving surveillance of moving objects.