基于语义分割的城市航拍图像匹配定位方法  

Localization method by aerial image matching in urban environment based on semantic segmentation

作　　者：郝运 孟子阳[1] 艾嘉文 吴元清 HAO Yun;MENG Ziyang;AI Jiawen;WU Yuanqing(Department of Precision Instrument,Tsinghua University,Beijing 100084,China;School of Automation,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]清华大学精密仪器系,北京100084 [2]广东工业大学自动化学院,广东广州510006

出　　处：《华中科技大学学报（自然科学版）》2022年第11期79-84,共6页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家自然科学基金资助项目(U19B2029,61873140);北京市自然科学基金资助项目(JQ20013)。

摘　　要：针对无人机如何在无全球导航卫星系统(GNSS)或弱GNSS环境下确定自身的位置,以使无人机完成各项既定任务的问题,提出一种在城市环境中以无人机上搭载的俯视相机和高度计作为传感器,通过航拍图像和已知参考卫星图像进行配准以获得无人机绝对位置的方法.在配准之前,使用卷积神经网络对两幅待配准图像进行语义分割,提取两幅图像中的建筑物信息.与传统点特征匹配方法容易失败不同,利用两幅图像的建筑物排列信息配准,可以克服两幅图像可能存在的部分差异,如季节等因素的变化.本方法无须已知无人机的精确初始位置,可利用初始化算法在大面积范围内确定无人机的初始位置.在视觉里程计(VO)运行过程中利用最近点迭代(ICP)点云配准算法以消除漂移.最后,利用在谷歌地球中模拟的数据集,验证了本方法的有效性.To address the problem that how to obtain the position of unmanned aerial vehicles(UAVs) in global navigation satellite system(GNSS)-denied environment or weak GNSS environment to enable them accomplish various established tasks,a method to obtain the absolute position of UAV in urban environment was proposed,by using a downward-facing monocular camera and an altimeter on the UAV as sensors,and the pre-existing reference satellite imagery to align with aerial imagery.Before alignment,the convolutional neural network(CNN) was used to extract the building information of the aerial imagery and pre-existing reference satellite imagery through semantic segmentation. In contrast to the traditional point feature matching method,by utilizing the building arrangement information to align the two imageries,and therefore the influence of the difference between the two imageries could be eliminated,such as season changes and so on.The proposed method could be performed without knowing the exact initial position of UAV,and the initialization algorithm could be used to estimate the initial position of UAV in the range of a large area.The iterative closest point(ICP) algorithm was used to eliminate the drift in the visual odometry(VO) process.Finally,experiment on the dataset simulated in Google Earth verified the effectiveness of the proposed method.

关 键 词：语义分割 图像匹配 视觉里程计 最近点迭代算法 无人机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]