Radar and AIS Track Association Integrated Track and Scene Features Through Deep Learning

The navigation accuracy of the ship can be improved by associating the output tracks of radar with that of the automatic identification system (AIS). The traditional association method only uses the track features (including position, speed, and heading angle) and has low association accuracy and weak scene adaptability. We integrate the track and scene features through deep learning to associate the radar and AIS output tracks. First, we employ the autoencoder (AE) to reduce the noise of the original track data. The statistical distances of the two tracks in position, speed, and heading are obtained as the new track data. We map the track data to the high-dimensional space to extract the track features through the fully connected and dropout layers. Then, we reconstruct the scene data using the Euclidean distances of all tracks in position. The scene data are input into the 3-D convolutional and max-pooling layers to extract the scene features. Finally, we fuse the track and scene features in the channel dimension by the concatenated layer and use the fully connected layer to estimate the probability of two tracks coming from the same target. The simulation results demonstrate that the proposed method has better scene adaptability and higher association accuracy than the traditional methods.