Unmanned Aerial Vehicle Image Localization Technology Based On Building Semantic Features

In recent years,unmanned aerial vehicles(UAVs)have been widely used in earth mapping,public security,precision strike and other military and civilian fields.Locating the content of UAV images is the basis of various typical applications.It is not only the internal task of UAV image understanding,but also the key to the subsequent completion of target attack,target tracking and others.Due to the low precision of the positioning devices,the drift and multi-path of UAV,the localization error is large with traditional aerial image localization method.Also,traditional localization method has complex process and need to laid control points outside,which is difficult to meet the practical needs of timely UAV image localization such as emergency support.In view of this,this paper combines UAV POS information and image,transforms the UAV image localization problem into the matching problem between UAV image and satellite remote sensing image with geographic location,and studies the efficient and accurate method,which has important theoretical significance and application value.The main works of this paper are:1.A building extraction network named EEMS-Unet is proposed to extract building semantic features from UAV images.In remote sensing image,building scale is different and its edge is not clear,which leads to poor building extraction effect,especially for irregular buildings.In this paper,EEMS-Unet(Enhanced Edge and Multi-Scale building extraction Unet)is proposed based on the characteristics of buildings.The innovations are designing multi-scale expansion convolution kernel module and introducing structural similarity loss function.Experiments show that the EEMS-Unet network proposed in this paper can improve the building extraction performance of building edges,reduce the loss of details and solve the problem of incomplete semantic extraction of irregular buildings in some extent.This network achieves better performance than others: 9.7%?11% higher than Unet on IOU and 7.8%?4.7% higher than Unet on F1-score on WHU-045 satellite images dataset and Hunan Shaoyang UAV images dataset.2.A semantic feature registration method named WASICP is proposed to align building semantic features from UAV image and satellite image.The building extraction result of UAV image is not completely accurate and there are many noise points and abnormal points.The same building object on UAV image and heterogeneous satellite image has different scale and is not consistent.This paper puts forward an improved registration algorithm named WASICP(Weighted Anisotropic Scaled Iterative Closest Point).This algorithm is applicable for building semantic features matching between UAV image and heterogeneous satellite image,and improves the registration performance by eliminating wrong matching points,weighting correct matching points by Euclidean distance,so as to overcome the failure of traditional feature matching methods.The experimental results show that WASICP is more effective than a series of ICP methods on public standard dataset and building semantic features matching data.3.A novel and fast two-stage UAV image localization method is proposed based on coarse localization by photogrammetry and fine registration by computer vision,which achieves the second-level and ground-level localization of UAV images.Based on the advantages of photogrammetry and computer vision method,a two-stage UAV image localization method is proposed in this paper.This method firstly uses photogrammetry to conduct rough position,and its localization error depends on the POS information accuracy carried by drone.Then,the building semantic features of UAV images are extracted by EEMS-Unet,and those corresponding building objects in satellite image is aligned by WASICP.This method does not need to set control points outside,and can locate of UAV images at the second-level and ground-level only by original UAV image and POS information.