YOLOv8-P2-CBAM-NWD: High-Precision Small-Target Drone Detection in Complex Visual Airspace via Shallow Attention Routing and Normalized Wasserstein Distance

Abstract

Optical ground-to-air detection of small unmanned aerial vehicles (UAVs) supports perimeter security and airspace safety, but a distant UAV may occupy fewer than 16 × 16 pixels, which is 0.0625% of a 640 × 640 detector input. Strided downsampling makes such objects harder to detect, and Intersection over Union (IoU) based box losses are highly sensitive to pixel-level shifts. This paper evaluates YOLOv8-P2-CBAM-NWD, a nano-scale version of YOLOv8 with a stride-4 P2 detection head, a Convolutional Block Attention Module (CBAM) on the shallow P2 branch, and a box loss that combines Complete IoU (CIoU) with the Normalized Gaussian Wasserstein Distance (NWD). Each model was trained only once on 800 single-class drone images and compared against a YOLOv8n baseline. On 200 validation images, the proposed model reached a mean average precision at an IoU of 0.5 (mAP@0.5) of 89.8%, which is below the 92.5% of the baseline. On a sentry test set of 2,500 frames (1,800 drone frames and 700 cloud or clear-sky frames) at a confidence threshold of 0.1, it increased frame-level precision from 96.94% to 98.35% and specificity from 92.14% to 95.86%, and it decreased false alarms from 78.57 to 41.43 per 1,000 negative frames, a 47.3% relative reduction. Frame-level recall went down from 96.78% to 96.11%. Thus, the main advantage is fewer false alarms rather than higher object-level accuracy. Since the threshold was selected on the test set and the frames were not grouped by video, these results are preliminary.