Unsupervised Trajectory Anomaly Detection for Situation Awareness in Maritime Navigation

Situation awareness is essential in conducting effective collision avoidance in potential ship encounter situations. It has been shown that data driven trajectory prediction techniques, utilizing historical AIS data, have the potential to aid in providing such awareness. However, such data driven techniques will not perform well for unusual ship behavior, i.e. anomalous trajectories. Additionally, such anomalies in the dataset can corrupt the predictions. In this study, an unsupervised approach to anomaly detection is presented to aid such trajectory predictions. Gaussian Mixture Models are used to cluster trajectories, such that clusters of both normal and anomalous trajectories are discovered. Further, anomalies are discovered within clusters of normal behavior. Novel trajectories can then also be evaluated based on a parametric description of the historical ship traffic. The approach is shown to be effective in detecting anomalies relevant in such a trajectory prediction scheme.