Track-to-track association from diverse source ship trajectory based on an improved graph neural network

With the development of maritime surveillance technologies, large volumes of ship trajectory data have been collected through various monitoring methods. The accuracy of targets tracking can be significantly improved by associating these trajectories, particularly in cases of partial equipment monitoring failures. Different acquisition approaches and data standards of marine monitoring methods lead to differences in the quality and precision of ship trajectory data from diverse sources, which significantly affects the accuracy of track-to-track association. According to AIS and Radar trajectory data, a novel multi-graph neural network combined with Cross-Attention (CA-MGNN) is proposed to realize the association with precision bias. The MGNN module is employed to convert AIS and Radar trajectory points into corresponding features, which are extracted from multiple spatiotemporal dimensions. The interactive features are then constructed by Cross-Attention mechanism, thereby achieving the association based on trajectory features. Experimental results demonstrate that the proposed method achieves high efficiency with association accuracy of 92.61% and inference time of 83.4 (ms), thereby providing more accurate tracking information to meet the demands of maritime awareness application. • Although many articles have extensively researched on ship data analysis,concentered on AIS and radar data to correlate and track the trajectory of individual ships. • An advanced method is designed to construct ship trajectory features through Multi-Graph Neural Network, effectively solving the problem of spatiotemporal asymmetry in multi-source ship data. • A Tracks-to-Tracks association module based on trajectory features is proposed to reduce the impact of accuracy differences in multi-source ship data. The excellent performance of the proposed method has been demonstrated through convincing numerical analyses.