Improved sequence-to-sequence ship trajectory prediction based on AIS

Abstract

Due to the high density of ships and the complex traffic environment in coastal areas, developing ship trajectory prediction methods is an imperative task for effective collision avoidance. The performance of most previous methods is limited by the issues of regional discrepancy of Automatic Identification System (AIS) data and singularity in trajectory feature extraction. To address these two issues, this paper proposes an improved ship trajectory prediction framework based on AIS. The proposed framework mainly consists of two parts: (1) A trajectory data preprocessing module including the extraction of trajectories based on time and ship speed for imputing missing values based on which AIS data from different regions is normalized to keep the time interval consistency. (2) A sequence-to-sequence ship trajectory prediction model based on double CNNs called DCNN. The encoder of DCNN is composed of two parallel structures, namely the global temporal convolution and local temporal convolution, to extract both long-term and short-term dependencies of the ship trajectories. The decoder is composed of RNN and its variants which are used to generate the predicted ship trajectory sequence. The experimental results show that the proposed framework can effectively predict the ship's trajectory, so as to predict the ship encounter in different regional traffic modes in advance, and help the ship to actively avoid collision.