Along with the evolution of scientific knowledge, UAV aerial photography technology has become an important tool for acquiring information, and the research of UAV target detection will contribute to the decision-making and guidance of traffic management and diversion, with great prospects in the future field of intelligent transportation. This paper proposes an algorithm model of improved YOLOv5 (Drone-YOLO) for target detection based on the expansion of UAV image data by utilizing the improved cyclic generative adversarial network (CycleGAN). The Drone-YOLO algorithm model improves the adaptability to the size changes of UAV image data targets through the increase of detection branches, integration of multi-level information, and fusion of multi-scale features. Additionally, the introduction of the multi-scale attention mechanism increases the attention ability of the network on the target objects; the decoupling of classification and regression tasks promotes the detection accuracy, and the optimization of the loss function improves the efficiency of training. According to the experiments conducted on the drone image data set of VisDrone, the precision of the Drone-YOLO algorithm reaches 51.5%, an increase of 5.6% from the original model. Meanwhile, the recall rate was improved from 34.2% to 39.6%, mAP@.5 from 34.5% to 39.8%, and mAP@.5:.95 from 18.3% to 23.1%, all of which are enough to complete the detection task of target objects in UAV scenarios.
