Aiming at the UAV viewpoint target scale change is large, the existence of occlusion and other circumstances lead to the model detection accuracy is low, there are problems such as misdetection, omission detection and so on. In this paper, a UAV weak target detection method is proposed by improving the YOLOv5 target detection algorithm. The method introduces dynamic full-dimensional convolution, regression loss function and increased detection head, and conducts a large number of experiments on the Visdrone dataset. The experimental results show that the improved YOLOv5 improves the mean accuracy (mAP) on the Visdrone dataset by 14% and the mAP@0.5:0.95 value by 11%.