An efficient and accurate YOLO-based framework for air-to-air UAV detection

Robust Air-to-Air (A2A) detection of unmanned aerial vehicles (UAVs) is critical for the safety of low-altitude airspace. This paper proposes a lightweight yet highly accurate YOLO-based detector to address this challenge. Our framework integrates three novel modules: the ASF-P for enhanced multi-scale feature fusion, the AB-CGLU for attention-boosted feature representation, and the ADown for efficient downsampling. This design effectively tackles small UAV detection in complex backgrounds while maintaining a compact architecture. On the primary dataset, our model achieves state-of-the-art performance with 98.1% precision, 84.4% recall, 91.6% mAP@0.5, and 57.1% mAP@0.5:0.95, while requiring only 2.03M parameters and 4.7MB of storage. It outperforms YOLOv8n by a large margin (e.g., +12.6% mAP@0.5) and also surpasses heavier YOLO variants in accuracy. Critically, cross-dataset validation over three independent runs on challenging unseen benchmarks confirms its strong generalization capability. This work provides an efficient and reliable solution for real-time A2A UAV perception, with immediate potential for deployment on resource-constrained autonomous aerial platforms.