Research On Orthophoto Stitching Algorithm Of UAV Monocular Aerial Image

With the popularity of UAVs,its application in people’s daily life is more and more extensive.Aerial photography is one of the most important applications of UAVs at present.Using UAVs,we can easily and quickly obtain real-time,high-resolution and top-view pictures.In many application scenarios,aerial image stitching algorithms are required to stitch all the images together to perceive the entire shooting scene.According to the different generation methods of the stitched images,the aerial images stitching algorithm can be divided into the aerial images stitching algorithm based on image fusion and the aerial images stitching algorithm based on 3D reconstruction.This paper will start with these two algorithms,and propose some improvement schemes to solve the shortcomings of the existing stitching algorithms.In the aerial images stitching algorithm based on image fusion,an algorithm that uses GPS heuristics to search for images to be matched is proposed.Using GPS to exclude images that cannot have common viewing areas in advance can speed up the process of image matching? Using the rotation matrix to measure the motion of the camera can reduce the number of estimated parameters in the Bundle Adjustment and reduce the optimization difficulty of the Bundle Adjustment.Meanwhile,an algorithm to initialize the rotation matrix using GPS is proposed.The algorithm can effectively reduce the error accumulation phenomenon in massive aerial images stitching,and improve the effect of massive aerial images stitching.Experiments show that,compared with Pix4 D,the algorithm proposed in this paper can obtain the approximate stitching results as Pix4 D on the dataset of plain terrain.What’s more,it can stitch the image boundary and texture area well,and it can also reduce the impact of moving objects.In addition,the algorithm proposed in this paper is dozens of times faster than Pix4 D,and it can quickly stitch massive aerial images.In the aerial images stitching algorithm based on 3D reconstruction,dozens or even hundreds of common-view images are required to generate a dense depth map when the MVS（Multi View Stereo）algorithm generates a depth map of a picture.This paper also proposes a MVS spatial regularization method,which can reasonably generate dense point clouds even when there are only a few common-view images,which can reduce the number of UAV shots.Experiments show that MVS with spatial regularization can reasonably generate denser point clouds.In the plain terrain data set,the number of point clouds generated is about 3 times that of the MVS algorithm without spatial regularization.The algorithms proposed in this paper can improve the effect,but still cannot satisfy the requirements of real-time.In addition,a better combination of aerial images stitching and machine learning is also a direction that can greatly improve the stitching algorithm.