Loitering behavior detection and classification of vessel movements based on trajectory shape and Convolutional Neural Networks

Abstract

To improve the ability of ship behavior detection in maritime safety surveillance, a new ship behavior of frequent turning within a certain spatial range, called ship loitering, was found under the ship behavior framework of "anchored-off, straight-sailing, turning". The concept of trajectory redundancy is defined as the quantitative representation of trajectory characteristics of ship loitering. A multi-scale sliding window-based ship loitering detection method is designed to process ship trajectories with different spatial ranges and time duration. Further, a shape recognition model based on a convolutional neural network is constructed and trained to identify four typical shapes of loitering trajectories. To verify the effectiveness of the proposed method, experiments were conducted using the AIS trajectory data collected from the sea off the Strait of Juan de Fuca in the eastern North Pacific Ocean in 2017 and validated against the manual labeling results. The results illustrate that six general categories of ships produce loitering behaviors with different intentions. The characteristics of the spatial-temporal distribution of ship loitering behaviors can be represented as different shapes. The average precision of ship loitering detection reaches 96.5% and the average accuracy of four typical loitering trajectory shapes reaches 86.2%.