Quantitative assessment of anomaly detection algorithms in annotated datasets from the maritime domain

The early detection of anomalies is an important part of a support system to aid human operators in surveillance tasks. Normally, such an operator is confronted with the overwhelming task to identify important events in a huge amount of incoming data. In order to strengthen their situation awareness, the human decision maker needs an support system, to focus on the most important events. Therefore, the detection of anomalies especially in the maritime domain is investigated in this work. An anomaly is a deviation from the normal behavior shown by the majority of actors in the investigated environment. Thus, algorithms to detect these deviations are analyzed and compared with each other by using different metrics. The two algorithms used in the evaluation are the Kernel Density Estimation and the Gaussian Mixture Model. Compared to other works in this domain, the dataset used in the evaluation is annotated and non-simulative.