Abstract
The intricate target sizes and fuzzy edges in Unmanned Aerial Vehicle(UAV) aerial images pose a great challenge for target detection. Therefore, investigating multiple recognition methods is crucial to solving the target detection problem of UAVs in challenging environments. In this study, the VisDrone2019 dataset is used, which contains high-resolution images under a variety of weather conditions and covers a number of challenging scenarios such as urban, suburban, and rural areas. In this paper, we propose a UAV aerial image detection and recognition method DAN-YOLOv8 based on the improved YOLOv8 network, which improves the detection performance of the model on UAV aerial images as well as the bounding box regression performance by replacing the convolutional part of the C2f module and redesigning the Neck part of the Backbone, as well as by introducing the Normalized Wasserstein Distance(NWD) loss function. After analyzing the UAV aerial images, it is observed that the detection model exhibits excellent environment perception and target recognition capabilities, which can simulate the intelligent features presented by humans when observing and analyzing images. The experimental results show that compared with other models, the network model proposed in this paper has a smaller number of parameters but higher accuracy, which significantly improves the detection performance and feature fusion ability of the model and enhances the detection efficiency and bounding-box regression performance for UAV aerial images. The results provide significant value and guidance for further development and application in the field of UAV target detection in challenging environments.