SPAE-YOLOv8 for Onboard Real-Time Perception: Lightweight Small UAV Detection from Air-to-Air Perspectives

The increasing use of UAVs has raised concerns regarding public safety and airspace security. To address air-to-air micro-UAV detection with cluttered backgrounds, tiny targets, and diverse viewing angles, this paper develops SPAE-YOLOv8, a lightweight detector based on YOLOv8n. SPAE consists of four core designs: SIoU loss, P2 shallow feature layer, ADown adaptive downsampling, and Efficient_UAVDet lightweight detection head. These modules improve small-target representation and reduce model size. In this paper, lightweight refers to the combination of parameter count, storage volume and inference speed. On the Det-Fly dataset, the proposed method achieves an mAP@0.5 of 0.922, outperforming YOLOv8n by 7.2 percentage points while reducing total parameters by 30%. We conduct independent training and testing on the DUT Anti-UAV dataset and obtain an mAP@0.5 of 0.906. Cross-dataset testing is further carried out on the more challenging Anti-UAV300 dataset without additional fine-tuning to verify the generalization performance of the model. In real-world onboard deployment, the model is implemented on an Intel NUC11TNHi7 embedded UAV platform with OpenVINO acceleration and achieves 43.9 FPS at a resolution of 640×640, satisfying real-time inference requirements. The ablation results demonstrate the contribution of the proposed modules, providing an efficient lightweight solution for airborne monitoring and civil airspace security.