Research On Stiching Algorithm Of UAV Aerial Images

UAV aerial images provide extensive information for land monitoring,disaster assessment,remote sensing mapping,and crop growth monitoring.To obtain a highresolution image of a large area,multiple aerial images need to be stitched into a panorama image.Although great progress has been made in the research on image stitching of UAV aerial photography,most algorithms are limited when performing bundle adjustment for registration parameters of multiple images and extracting features from images with similar texture.The study of these problems has practical application value,and is beneficial to the research development in the field of image stitching.To enhance the accuracy of image stitching,this thesis aims build a UAV aerial image stitching system and focuses on establishing a model with smaller stitching errors.This thesis explores from the perspective of template-based registration and feature-based registration respectively,which effectively improves the stitching performance.The main work of this thesis can be described as follows:On the one hand,in single viewpoint mode,most template-based registration methods have poor performance on bundle adjustment for registration parameters of multiple images.Therefore,this thesis proposes a two-level error dispersion method based on binary tree to globally optimize the parameters after registration.The effectiveness of the method is validated through comparative experiments.On the other hand,in the multiple viewpoints mode,the original HomographyNet regression network is improved,and data augmentation is used to improve the distribution of dataset,which achieve image registration with complex transformation relationship in multi-viewpoint aerial photography mode.Since traditional methods have difficulty in extracting complete different features when registering images with similar textures,the registration has a large error,even fails.Our method has advantages in solving this problem.The comparison experiments validate the superiority of our method.