Density-Based Anomaly Detection in the Maritime Domain

Abstract

Detecting anomalies in the maritime domain is a complex task. The complexity is due to the many information sources in the maritime domain and the knowledge required to interpret the information. Operators that monitor vessel movements may be supported by systems that help them to identify abnormal behaviors and events. The uncertain nature of some aspects of the maritime domain (e.g., weather predictions), suggests the use of statistical computer methods, rather than deterministic rule-based methods. In these methods, maritime objects and events (e.g., vessels and vessel behaviors) are represented as points in a (potentially high-dimensional) space in which the dissimilarities of pairs of points are represented by distances. Anomalies manifest themselves as points that are distant from all other points. We present the Stochastic Outlier Selection (SOS) method that takes an unlabeled set of points and automatically selects the outliers. In a comparative evaluation involving a wide range of complex (non-maritime) tasks, the SOS method is shown to outperform state-of-the-art outlier detection methods. We conclude that the SOS method is suitable to be applied to complex tasks.