Small Object Detection of UAV Based on NF-YOLOv5 for Large-scale Activities

Large-scale activities generate additional traffic demand on the basis of daily traffic flow in a fixed period of time. In order to ensure the smooth progress of activities and the safety of participants, unmanned aerial vehicles (UAV) were widely used in large-scale activities. Aerial images of UAV have the characteristics of high density of small objects, target occlusion and illumination change. At present, many researchers have studied aerial object detection, but challenges still exist, such as unsatisfactory detection accuracy and efficiency. To address the problems, we propose NF-YOLOv5 model. First, the backbone incorporates the NFN et network for normalization-free training, accommodating larger batch sizes and enhanced data augmentation. Next, Convolutional Block Attention Module(CBAM) is added at the end of the backbone network to enhance channel and spatial domains and reducing computational costs. Subsequently, new feature fusion layers and prediction heads are added at shallow layers for small object detection. Finally, the EIoU is used to simplify the overlap between prediction frame and real frame in object detection algorithm. Experimental results show that the improved network increases mAP50 by 7.9% and mAP50-95 by 5.4% on the VisDrone-2019-DET dataset.