Research On Fast Stitching Methods Of UAV Aerial Images Based On Local Features

In recent years,unmanned aerial vehicles(UAV)have been widely used in various fields of life.In UAV aerial photography,image stitching technology can be used to obtain highresolution images with a large field of view.At the same time,in real-time applications such as disaster relief and emergency,the overall situation of the disaster area must be obtained quickly in order to formulate a rescue plan.Therefore,it is necessary to study fast and accurate UAV aerial image stitching methods.At present,image stitching technologies are mostly based on the local features of the image,i.e.,features of the highly distinguished areas in the image,which has good robustness to the changes in the image,such as illumination and rotation.This thesis mainly researches the aerial image stitching methods of unmanned aerial vehicles based on local features from two aspects: image registration and image fusion.The main contributions are as follows:(1)With respect to image registration,this thesis proposes an improved ORB algorithm,which introduces scale space and uses Hessian detection operator for feature extraction;at the same time,combines the Grid-based Motion Statistics(GMS)algorithm with the Progressive Sample Consensus(PROSAC)method to purify feature matching points.Simulation results show that compared with the ORB algorithm,the improved algorithm increases the accuracy of feature point matching by about 3%,and reduces the registration error by about 22.72%.(2)With respect to image registration,this thesis also proposes a stitching method of multiple UAV aerial images combining position and attitude information of UAV with the improved ORB algorithm.This method first uses the improved ORB algorithm to obtain the transformation relationship between the images,then corrects the geographic location coordinates of each image according to the transformation relationship,and finally performs the image stitching according to the location coordinates.Experiments show that,compared with the improved ORB algorithm,the new method can eliminate the cumulative error of the stitching of multiple aerial images,and reduces the registration error by about 26.82%.(3)With respect to image fusion,this thesis proposes a fusion strategy based on category for multiple UAV aerial images.The root mean square error of the color of the overlapping area between registered images is used to determine the fusion method of this area,which effectively solves the problem of requiring different fusion methods due to the differences in various stitching points when stitching multiple images.Experiments show that when multiple images are fused,the average gradient of this method is increased by about 11.30% compared with the weighted average fusion method of fade-in and fade-out,and the time consumed is reduced by about 18.89% compared with the best stitch fusion method.Therefore,the new fusion method achieves a good balance between fusion speed and fusion performance.