Learning for Prediction of Maritime Collision Avoidance Behavior from AIS Network

With the rapid increase in global maritime shipping, there is a great demand for the technology of maritime traffic monitoring to detect inappropriate encountering interaction between ships and prevent ship collision accidents. The Automatic Identification System (AIS) network makes it possible to collect a large volume of maritime traffic data and investigate the collision avoidance behavior of real-world ships. Most collision avoidance systems are based on expert systems and simulations based on the International Regulations for Preventing Collisions at Sea (COLREGs). Those regulations outline the general principles underlying collision avoidance; however, they do not provide specific guidance and fail to account for the complexity of many real-world situations. Furthermore, guidance systems coordinating the movement of a ship must have the capacity to predict the movement behavior of all ships involved in potential encounter situations, and do so as early as possible for anti-collision reaction. Our objective in this study was to model the collision avoidance behaviors of human operators in order to formulate a set of realistic trajectory predictions for encountering near collision scenarios. By machine learning approach, the proposed framework is able to learn a model of interaction movement behavior from collected AIS historical traffic data involving near collision situations and then generate a set of predicted trajectories while ships encountering. The proposed model eliminates the need for a priori information related to environmental conditions and the rules governing encounter situations. The resulting projections can be used to suggest anti-collision paths for navigators or to guide the selection of collision-free paths for maritime autonomous surface ships.