Data-driven detection and context-based classification of maritime anomalies

Abstract

Discovering anomalies at sea is one of the critical tasks of Maritime Situational Awareness (MSA) activities and an important enabler for maritime security operations. This paper proposes a data-driven approach to anomaly detection, highlighting challenges specific to the maritime domain. This work builds on unsupervised learning techniques which provide models for normal traffic behaviour. A methodology to associate tracks to the derived traffic model is then presented. This is done by the pre-extraction of contextual information as the baseline patterns of life (i.e., routes) in the area under investigation. In addition to a brief description of the approach to derive the routes, their characterization and representation is presented in support of exploitable knowledge to classify anomalies. A hierarchical reasoning is proposed where new tracks are first associated to existing routes based on their positional information only and "off-route" vessels are detected. Then, for on-route vessels further anomalies are detected such as "speed anomaly" or "heading anomaly". The algorithm is illustrated and assessed on a real-world dataset supplemented with synthetic abnormal tracks.