AbstractsComputer Science

The operators of a maritime surveillance system are hardpressed to make complete use of the near real-time informationflow available today. To assist them in this matterthere has been an increasing amount of interest in automated systems for the detection of anomalous trajectories.Specifically, it has been proposed that the framework of conformal anomaly detection can be used, as it provides the key property of a well-tuned alarm rate. However, inorder to get an acceptable precision there is a need to carefully tailor the nonconformity measure used to determine if a trajectory is anomalous. This also applies to the features that are used by the measure. To contribute to a better understandingof what features are feasible and how the choice of feature space relates to the types of anomalies that can be found we have evaluated a number of features on real maritime trajectory data with simulated anomalies. It isfound that none of the tested feature spaces was best for detecting all anomaly types in the test set. While one feature space might be best for detecting one kind of anomaly,another feature space might be better for other anomalies.There are indications that the best possible non conformity measure should capture both absolute anomalies, such asan anomalous position, as well as relative anomalies, such as strange turns or stops.