An Efficient Framework for Track-to-Track Association via Graph Neural Networks in Large-Scale Dense Scenarios

Track-to-track association (T2TA) plays a crucial role in multisensor data fusion systems, serving as an essential prerequisite for track-to-track fusion. In this article, we propose a novel T2TA framework that efficiently maps large-scale dense tracks ( ${\sim }{10}^{{3}}$ in quantity) to an optimal association matrix, addressing the severe performance degradation (e.g., accuracy and efficiency) faced by existing methods. First, the temporal and spatial features (local features) of individual tracks are extracted using an LSTM autoencoder. Second, these local features are transformed into graph representations, where node aggregation is performed through an attention-based graph neural network (GNN) to capture intertrack coupling features (global features). Third, adaptive feature fusion is performed to construct the association cost matrix using cosine similarity. The optimal association matrix is then derived by solving the 2-D assignment problem, preserving physical constraints, with a differentiable Sinkhorn algorithm that supports backpropagation for network training. Finally, comparative analyses with three representative baselines reveal significant improvements in both T2TA accuracy (improved by 35% over statistical distance-based and reference topology-based methods) and computational efficiency (exceeds two orders of magnitude faster than an advanced learning-based method) of the proposed approach. Additionally, extensive experiments under varying parameter settings (e.g., the number of tracks, track densities, and sensor characteristics) demonstrate the robustness of the proposed approach, particularly in large-scale dense scenarios.