激光与光电子学进展, 2017, 54 (11): 111002, 网络出版: 2017-11-17
基于YOLO v2的无人机航拍图像定位研究 下载: 2073次
Aerial Image Location of Unmanned Aerial Vehicle Based on YOLO v2
图像处理 卷积神经网络 目标检测 图像定位 image processing convolutional neural networks YOLO v2 YOLO v2 object detection image location
摘要
为了保证定位的速度和准确率,采用2016年在目标检测领域取得最佳检测效果的YOLO v2网络制作了以明显特征的地物作为目标区域的目标检测数据集。通过目标框维度聚类、分类网络预训练、多尺度检测训练及更改候选框的筛选规则等方法改进YOLO v2网络,使其更好地适应定位任务。能够将无人机实时获取的航拍图像定位到目标区域,并通过投影关系进行坐标转换得到无人机的经纬度。结果表明:该方法效果较为理想,在航拍图像的目标区域检测任务中检测网络的平均准确率提高到79.5%;在包含目标区域的航拍图像中,经模拟飞行的仿真实验验证,其网络定位准确率大于84%。
Abstract
In order to ensure the speed and accuracy rate of location, the YOLO v2 network with the best detection effect in the field of object detection in 2016 is used to make the target detection data sets with the obvious features of surface features as the object area. Through the dimension clustering of object box, classified network pre-training, multi-scale detection training, change the candidate box filtering rules and other methods, the YOLO v2 network is improved, and it can better adapt to the location task. The network is able to locate the object area in the aerial image acquired from the unmanned aerial vehicle in real time. And the latitude and longitude of unmanned aerial vehicle are obtained by the projection relationship and coordinate transformation. The experimental results show that the proposed method can achieve better effect, and the average accuracy rate of the detection network increases to 79.5% in the object area detection task of the aerial image. It is verified by simulation experiment of simulated flight, the accuracy rate of the network location is over 84% in the aerial image that contains the object area.
魏湧明, 全吉成, 侯宇青阳. 基于YOLO v2的无人机航拍图像定位研究[J]. 激光与光电子学进展, 2017, 54(11): 111002. Wei Yongming, Quan Jicheng, Hou Yuqingyang. Aerial Image Location of Unmanned Aerial Vehicle Based on YOLO v2[J]. Laser & Optoelectronics Progress, 2017, 54(11): 111002.