AirVision: a fast and efficient small target detection method for unmanned aerial vehicles using You Only Look Once

Unmanned aerial vehicle (UAV)-based object detection has gained increasing traction in both civil and military applications. However, detecting small targets in aerial imagery remains difficult due to significant scale variations, dense distributions, overlapping instances, and complex backgrounds. To address these challenges, we propose AirVision-YOLOv10n, a lightweight model optimized specifically for small target detection in UAV imagery. Our key innovations include (1) a small-target feature pyramid network that integrates spatial depth-transformed convolution within an enhanced cross stage partial omni-kernel module backbone, significantly improving fine-grained feature representation; (2) a task-specific adaptive head incorporating group normalization, which enhances detection head efficiency through parameter sharing and adaptive multi-scale feature processing; and (3) strategic improvements, including deformable convolutions for flexible feature alignment, the content-aware reassembly of features module for content-aware feature upsampling, and WIoUv2 loss for refined bounding box regression, particularly effective for small and densely packed objects. Evaluated on the VisDrone2019 dataset, AirVision-YOLOv10n achieves a 5.6-point mAP improvement over the baseline YOLOv10n, with minimal additional parameter cost. Consistent and substantial performance gains are also observed on the TinyPerson and AI-TOD datasets, validating the model's robustness and generalization.