A Plot-to-Track Association Framework Based on Graph Representation Learning for Compact HFSWR

To address plot-to-track association ambiguities caused by low spatial resolution and measurement uncertainties of compact high-frequency surface wave radar (HFSWR) operating in complex maritime environments, a hybrid framework integrating a target state prediction method based on Transformer model and a plot-to-track association method using graph-based feature representation is proposed. In the target state prediction stage, the Transformer model employs self-attention mechanisms to capture long-term temporal dependencies in historical track data to improve the prediction accuracy of the plot-to-track association gate center. In the plot-to-track association stage, a graph autoencoder (GAE) is used to extract graph structure features of candidate measured plots from their corresponding range-Doppler (R-D) spectra. The GAE encodes the diffusion patterns and spatial correlations of radar echoes on the R-D spectrum into discriminative graph feature vectors. The graph feature vectors, which can enhance both inter-target discrimination and target-clutter differentiation, serve as the basis for association similarity computation. Experimental results demonstrate that the proposed framework significantly reduces target state prediction errors and suppresses false plot-to-track associations, achieving an average tracking time on target of 135.6 minutes, which is improved by 36% compared with the baseline method (motion model-based target state prediction combined with nearest neighbor data association using kinematic parameters).