Improved YOLOv5 for object detection in UAV aerial images

UAV aerial image object detection is of great significance for intelligent target identification and tracking, but the target under the UAV viewpoint is subject to large changes in target scale due to the influence of light, and there are cases of occlusion, low target resolution, etc., which lead to low model detection accuracy, misdetection, leakage and other problems. To address the above problems, an improved object detection method for UAV aerial images is proposed based on the YOLOv5. The method introduces Space-to-depth Convolution (SPD-Conv), Normalization-based Attention Module (NAM) and regression loss function, and conducts a large number of experiments on Visdrone2019 dataset. The experimental results show that the improved YOLOv5 algorithm improves the mean accuracy percentage (mAP) by 6.1% and the mAP@0.5:0.95 by 5%.