基于改进YOLOv3无人机图像目标检测

Target Detection Based on Improved YOLOv3 UAV Image

无人机航拍图像中的物体通常很小,边界模糊,加上复杂的背景和不断变化的照明条件,所以YOLOv3算法的检测精度相对较低.因此,构造四级BiFPN,既可以对融合的输出特征作出平等的贡献,又可以使得Neck部分变为FPN+PAnet结构,同时充分利用底层与高层特征融合信息,以提高特征提取和目标检测性能.添加更小的检测层,可以使得模型拥有四个先验框,有更高的概率出现对于目标物体有良好匹配度的先验框,使得模型更容易学习.实验结果表明,提出的无人机图像目标检测模型(YOLOv3_Drone)在VisDrone-2019的mAP比YOLOv3算法提高了3.83%,证明了该方法的有效性.

Objects in UAV aerial images are usually very small,with blurry boundaries,complex backgrounds and changing lighting conditions,so the detection accuracy of YOLOv3 algorithm is relatively low.Therefore,constructing a four level BiFPN can not only make equal contributions to the fused output features,but also make the Neck part become an FPN+PAnet structure;and meanwhile,it makes full use of the fusion information of low-level and highlevel features to improve the performance of feature extraction and target detection.By adding a smaller detection layer,the model can achieve four priori boxes,and further,there is a higher probability that a priori box with good matching degree for the target object will appear,which will make the model more easier to learn.The experimental results show that the proposed UAV image target detection model(YOLOv3_Drone)in VisDrone-2019's mAP is 3.83%higher than that of YOLOv3 algorithm,which proves the effectiveness of this method.