Long-term Vessel Motion Predication by Modeling Trajectory Patterns with AIS Data

Abstract

It is of critical importance for vessels to detect potentially hazardous situations as early as possible. Therefore, recognition and prediction of vessel motions require effective representations for analysis and clustering of motion trajectories. While short-term motion prediction of moving object is largely achievable, long-term prediction is more useful given the restricted maneuverability of vessels since a vessel cannot abruptly stop, turn or reverse as a land vehicle does. To this end, we propose in this study a long-term vessel motion prediction approach based on a combined trajectory classification and long short-term memory (LSTM) networks framework. As a measure for the similarity between trajectories, we introduce the longest common subsequence (LCS) algorithm to define trajectory similarity when making DBSCAN clustering. The grouped trajectories representing various motion patterns are further modeled via LSTM networks, in which vessel trajectory data are formed by regressing relative motion against current position and then the iterative prediction is applied for long-term prediction. We use the proposed approach for classifying and predicting motions in vessel traffic monitoring domains and test on real AIS data. Experiments show the benefit of this approach for long-term motion predication where parametric models such as Kalman Filters would perform poorly.