| Abstract [eng] |
This paper develops and assesses a comprehensive framework for collision risk using trajectory prediction regions and clusters. Using month’s worth of Automatic Identification System (AIS) data from the Danish Maritime Authority, RNN based models were created to predict the further trajectory of marine vessels. Ten noised predictions were utilized to create prediction regions. To evaluate collision risk, we implemented both boundary-based methods (ellipsoidal point intersection, ellipsoidal area intersection, and Z-score bounding box area intersection) and non-boundary-based methods (Gaussian Mixture Model (GMM) area intersection and a novel silhouette value-based score transformed via root functions). Results indicated that boundary-based methods performed poorly, particularly due to the "pass-by problem," where non-simultaneous region intersections fail to capture collisions effectively. In contrast, non-boundary-based methods, particularly GMM, demonstrated superior performance due to their continuous nature. GMM detected collisions with high accuracy, including identifying a real collision scenario up to 20 minutes prior to the event. These findings highlight the limitations of boundary-based metrics and underscore the potential of continuous, probabilistic approaches for improving maritime safety. |
