Hybrid-driven vessel trajectory prediction based on uncertainty fusion

Abstract

Predicting the trajectories of maritime vessels is significant for making early warning of potential collisions to reduce the maritime accident probability. As we know, the performance of most of approaches depends seriously on the utilized dataset or model. In this work, a novel hybrid-driven approach is proposed for maritime vessel trajectory prediction. The hybrid-driven approach is achieved by the uncertainty fusion of a data-driven predictor and vessel motion-based estimation. The data-driven predictor developed with a Long Short-Term Memory (LSTM) network has been trained by our dataset and has the ability to calculate the trajectory and uncertainty for the future moment. Then, uncertainty fusion is achieved to fuse the output of the data-driven predictor with the vessel motion estimation. The predicted trajectory sequence is more accurate, and the accompanying uncertainties can reflect the reliability of the hybrid-driven predictor. In addition, vessel trajectories from original Automatic Identification System (AIS) data are extracted for training and evaluating the proposed hybrid-driven predictor. Quantitative experiments and discussion are given in the end, and it is illustrated that the hybrid-driven predictions are practical for collision avoidance in actual maritime scenarios.