Anomaly detection method for ship trajectory based on stay region mining

Abstract

To address the issue of detecting anomalous sub-trajectories, such as circling and lane-changing behavior, an anomaly detection method for ship trajectory based on stay region mining (TADS) is proposed. Firstly, stay regions are mined using kernel density analysis based on stay points, which are obtained by combining heading deviation and heading zones. Secondly, sub-trajectories are generated by dividing ship trajectories based on stay regions, forming independent sub-trajectory sets both within and between different regions. And then, the similarity matrix between sub-trajectory features within each sub-trajectory set is constructed using edit distance. Finally, an adaptive hierarchical clustering algorithm is applied to detect anomalous sub-trajectories based on the similarity matrix. The AIS data from Lake Superior and Lake Huron are used as experimental samples to test TADS. The results indicate that TADS identified 11 and 7 stay regions in Lake Superior and Lake Huron, respectively, with an average trajectory anomaly detection accuracy of 91.27 $\%$, a precision of 94.65 $\%$, a recall of 94.72 $\%$, and an F-measure of 94.67 $\%$. Compared to similar methods, TADS exhibits better performance in detecting anomalous trajectories.