C2F-Net: Coarse-to-Fine Multidrone Collaborative Perception Network for Object Trajectory Prediction

Abstract

Multidrone collaborative perception network can forecast the motion trajectories of grounded objects by aggregating intragroup communication and interaction, exhibiting significant potential across various applications. Existing collaborative perception methods struggle to address the nonuniform spatial distribution of targets and the spatial heterogeneity of multisource perception information typical in remote sensing scenarios. To tackle these challenges, we propose a coarse-to-fine feature fusion network C2F-Net, utilizing coarse-grained information interaction to guide the fusion of fine-grained features. Our approach includes a selective coarse-to-fine feature collaboration module that estimates perception levels of specific areas based on bird's-eye-view features, selectively collaborates on sparse features according to complementary information principles, and achieves efficient spatial feature interaction and fusion. In addition, we employ a region-aware effectiveness enhancement module, leveraging the differences between swarm and individual perception as prior knowledge to guide regional perception level estimation, improving comprehensive environmental understanding. We also introduce a simulation dataset named CoD-Pred for multidrone collaborative trajectory prediction. Extensive experiments demonstrate that C2F-Net significantly improves the accuracy of multidrone collaborative trajectory prediction, which increases mIoU by 2.7% to 3.3% and VPQ by 1.0% to 9.1% under comparable information transmission conditions, offering an effective and efficient solution for multidrone collaborative perception.