SAD-YOLO: A Small Object Detector for Airport Optical Sensors Based on Improved YOLOv8

As a crucial technology for enhancing the autonomous perception capability of airport optical sensors, object detection has become a research focus. This article proposes a small object detector for airport optical sensors named SAD-YOLO to improve airport safety and assist airport traffic management. Firstly, our experiment finds that the YOLOv8x with the anchor box performs better in small object detection. Secondly, the ELAN-A module is designed based on YOLOv8x, which utilizes position, channel, and spatial attention methods to improve the neck and effectively enhance the feature fusion ability. Then, we introduce the Swin Transformer in the backbone and a SODL in the head to improve the small object feature extraction capability. Finally, we introduce the SIoU loss function to find the optimal model parameters for the proposed method. The airport traffic management system is equipped with high-performance computers. Therefore, this article focuses more on the accuracy and generalization of the proposed method. The comparison experiments show that the SAD-YOLO improves by 7.4% mAP on the ASS1 dataset compared to the official YOLOv8x. Compared to classical detectors, the proposed method obtains the optimal mAP on the generalizability experiments of MASD and ASS2 datasets, demonstrating that SAD-YOLO has excellent robustness and generalization.