AeroDet: lightweight multi-range context and global spatial refinement for dense aerial object detection
Year: 2026
Type: article
Abstract: Dense traffic imagery acquired by unmanned aerial vehicles (UAVs) contains compact, overlapping, scale-varying, and visually similar road users, making accurate detection difficult under tight model-complexity constraints. Existing UAV detectors often improve small-object representation through broader changes to feature pyramids, prediction heads, losses, or high-resolution processing, leaving open whether focused feature refinement can improve accuracy without expanding a compact reference model. We propose AeroDet, a lightweight YOLOv11n-derived detector that preserves the three-scale detection head and training objective while introducing Gated Multi-Range Aggregation (GMRAgg), Log-Scale Dilated Pooling (LogDilPool), and Mask-Gated Spatial Attention (MaskGatedAttn) for multi-range refinement, resolution-preserving contextual aggregation, and branch-restricted global spatial interaction, respectively. On the AeroBD validation set, AeroDet achieved 94.0% $${{{{\rm{mAP}}}}}_{50}$$ and 78.3% $${{{{\rm{mAP}}}}}_{50-95}$$ with 2.2 M parameters and 6.4 GFLOPs. On the held-out test set, it reached 93.8% $${{{{\rm{mAP}}}}}_{50}$$ and 77.5% $${{{{\rm{mAP}}}}}_{50-95}$$ , improving YOLOv11n by 2.5 and 4.4 percentage points while reducing parameters from 2.6 to 2.2 M. On the AeroBD validation set, small-object AP increased from 53.5 to 60.7%. Separately trained evaluations improved $${{{{\rm{mAP}}}}}_{50-95}$$ from 17.6 to 19.7% on VisDrone-DET2019 and from 23.0 to 25.9% on UAVDT. These findings show that targeted multi-range contextual and global spatial refinement improves dense aerial-object detection, particularly for small objects, within the computational budget of the compact reference detector.
Institutions: Donghua University, Beihang University, Fudan University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,990
OpenAlex ID: W7236863594
