Two-Stage Multi-Drone Multi-Target Association: Laplacian-Guided Feature Enhancement and SGMM-Based Error Filtering Approach

Cross-view target association is the foundation of multi-drone collaborative recognition. Existing methods often fail to fully exploit both the visual appearance and neighborhood spatial relationships of targets, making them vulnerable to large viewpoint variations and densely distributed similar objects, which frequently result in incorrect associations. To address this, we propose a two-stage cross-view target association framework, following the paradigm of “feature enhancement–error filtering” . In the feature enhancement stage, we implicitly model the neighborhood topological structure of each target and aggregate visual-semantic information from spatially adjacent targets via position-guided self-attention. Subsequently, for each target, we propose Laplacian-guided cross-attention to aggregate features of the target in another view whose appearance and spatial characteristics are most similar. In the error filtering stage, we propose a score-guided Gaussian Mixture Model (S-GMM) to filter error associations, and then find best homography matrix to achieve precise cross-view targets mapping. Our method demonstrates superior performance in scenarios with densely distributed and highly similar targets. Moreover, the homography matrix obtained in the second stage serves as a byproduct that can assist in handling target occlusions during subsequent collaborative recognition. The proposed method achieves promising results on the MDMT and M3Track dataset. The code will be released at https://github.com/bigdreaming11/LGSGMM.