A multi-graph aggregation network for robust multi-drone trajectory prediction

Multi-drone trajectory prediction is essential for enhancing the efficiency and safety of intelligent transportation systems. However, multi-drone collaborative perception still faces two major challenges. First, the perceptual quality of modality-specific information often degrades in complex environments, making it difficult to adaptively assign appropriate weights to different modalities. Second, existing methods largely overlook mutual learning among multi-view images, which limits the understanding of spatial structures and weakens the generalization capability of the models. To address these issues, this paper proposes a reliability-aware LiDAR-Camera fusion with adaptive multi-graph feature aggregation network (RFMANet) for multi-drone trajectory prediction. Specifically, we design a reliability-aware LiDAR-Camera fusion (RLCF) module that operates in the BEV space. We employ a dynamically gated multi-expert subnetwork to compute spatially adaptive modality weights, guiding cross-attention-based fine-grained feature fusion for subsequent graph construction. Furthermore, we introduce the adaptive multi-graph feature aggregation (AMGFA) module, which constructs collaborative feature maps from multi-view BEV representations. The Top-K sparse attention mechanism is employed to extract salient multi-scale features from different viewpoints, which are used to build structured multi-graph representations. Graph interactions capture both intra-view and cross-view dependencies, modeling geometric consistency and semantic complementarity. Experimental results on the Air-Co-Pred dataset show that RFMANet outperforms existing methods, demonstrating robustness and high prediction accuracy. We also simulate sensor degradation scenarios to further verify the robustness of the proposed method.