Adversarial Maritime Trajectory Prediction with Real-time Spatial-Temporal Mutual Influence

As the driving force of global trade, it is essential to ensure the safety and security of maritime transportation effectively. Due to human operators' physiology and experience limitations, it is unreliable to depend entirely on them for maritime surveillance. Fortunately, massive data collected by the Automatic Identification System (AIS) enables machineaided navigation and surveillance, based on which enormous studies on ship trajectory prediction have arisen for further maritime traffic monitoring/tracking, abnormal vessel detection, collision avoidance at sea, etc. However, previous works mainly predict vessel trajectories by modeling and grasping a single vessel's past navigation information based on well-cleaned AIS data, ignoring dynamic correlations caused by real-time ship interactions. In this work, we developed an Adversarial Maritime Trajectory Prediction with a Real-time Mutual Influence (MIAMTP) model to predict vessel trajectories in the same scene without complicated data pre-processing, which contains three main modules, namely trajectory embedding module (TEM), mutual influence module (MIM), adversarial prediction module (APM). Especially in MIM, by grasping the time dynamics, the proposed model can obtain real-time mutual influences of the vessels in order to predict the trajectories timely. The proposed MI-AMTP model was evaluated on one private maritime dataset collected from Southeast Asia and several publicly available AIS datasets. Experimental results denoted that the proposed model outperformed the state-of-the-art methods.