Visual Localization Of UAV Based On Reference Images

The drone has the advantages of being easy to carry,cheap,and flexible,and is widely used in aerial photogrammetry.However,since the image taken by the drone lacks pose information,it needs to be acquired through ineffective control points,GNSS or other post-processing methods.Using the visual localization algorithm to calculate the pose of the drone image is a more efficient method.This method firstly obtains the visual correspondence between the image of the drone to be located and the reference image with pose,and then uses the multi-view geometry to solve the pose of the drone image.Compared with the post-processing method,the visual localization can be fully automated,without manual interference.This paper fully studies the visual localization algorithm of UAV based on reference image and divides it into scene localization and pose localization.The scene localization acquires the approximate position of the drone,and the pose localization is calculated the exact position of the drone.The main work of this paper is as follows:First,before performing visual localization,we need to construct a scene map and a pose map using reference images.When constructing the scene map,the scene is divided into different blocks according to the scene range captured by the reference image.Then,on each block,the image block is intercepted from the reference image according to the visible condition of the block on the reference image.Finally,the visual features of the block are extracted from the image block using AlexNet-conv3 based on Image Net Dataset.When constructing the pose map,the SIFT feature points are extracted from the reference image.Then the feature matching is performed according to the pose information of the reference image,and the matching result is connected and cleaned by the Union-Find algorithm.Finally,we use triangulation to obtain the three-dimensional coordinates of the object point and map the object point to the SIFT descriptor one-to-one.Second,when performing visual localization of the drone,firstly perform scene localization.After extracting the visual features from the image using AlexNet-conv3,find the most similar visual features and corresponding reference images from the scene map according to the cos distance of the vector.Then according to the most similar reference image acquired from the scene localization,the corresponding object point and its corresponding SIFT descriptor are obtained from the pose map.Finally,the SIFT feature is extracted from the image and matches with the object points to obtain the 2D-3D correspondence,and the pose of the drone image can be obtained by using EPn P and Bundle Adjustment.Finally,we test the proposed algorithm on two different sets of data and results prove the feasibility and effectiveness of our algorithm.