Research On Key Technologies Of Incremental SFM Based On UAV Image

Image-based three-dimensional reconstruction is one of the important challenges in the field of photogrammetry and computer vision.The goal is to obtain useful geometric and semantic information from the images obtained through the camera.Image-based 3D reconstruction requires accurate calculation of the camera's internal and external orientation and other elements.In the field of photogrammetry,aerial triangulation technology is generally used to solve the pose of the image,but the initial parameters of the more accurate image pose and object point need to be improved in the air three process.Different from the aerial triangulation in traditional photogrammetry,the structure of motion recovery from computer vision has attracted extensive attention and research in academia due to the complete automation of the entire process.The motion recovery structure recovers the camera pose and the encrypted point coordinates in the scene from multiple images with a certain overlapping area and between different perspectives.In this paper,incremental SFM is used to sparsely reconstruct UAV aerial images,and research is conducted to improve the efficiency and robustness of SFM.The main contents of the paper are:1. The basic theoretical part of incremental SFM is described.The general camera aperture imaging model and distortion model are introduced.Introduced feature point extraction and matching,multi-view geometric foundation,including epipolar geometric model,triangulation and other knowledge,and deduced the key mathematical model.The theoretical derivation of the beam adjustment model is carried out,and the advantages and disadvantages of different solving methods are analyzed.2. In order to improve the efficiency of incremental SFM,this paper studies the time-consuming feature matching and beam adjustment.For images with POS data,pre-calculate the degree of overlap between the images to provide an initial matching list.For images without POS,use the pyramid matching method to determine the image matching relationship,and then use the pyramid matching method to guide the matching step by step to make the feature matching time.The complexity is reduced fromO(n~2)to the O(n),which can greatly shorten the time taken for matching.The use of Schur complement method to accelerate the adjustment of large-scale beam adjustment can greatly improve the efficiency of adjustment.The efficiency of this method when processing hundreds of images is higher than that of the commercial software Smart3D;compared with COLMAP,the efficiency of this method has increased significantly.3. In order to increase the robustness of incremental SFM,this paper studies how to eliminate the mismatch results and the new image addition strategy.In order to take into account the influence of the number and distribution of visible object points on the accuracy of image registration,this paper proposes a new method based on scoring to add new images to make the image addition more reasonable.Using RANSAC with geometric verification to eliminate mismatch points and dynamically determine the number of iterations based on the results of each iteration makes the incremental SFM system more robust.Multiple experiments show that the accuracy of the method in this paper the requirements of 1:500 mapping specifications.4. Aiming at the serious problem of accumulative error of self-checking beam method adjustment for UAV images in the strip area,a self-checking method for image classification based on the combination of KD tree and K-Means is proposed.First,build a KD tree based on the GPS position information of the image,determine the number of image sets to be divided and the seed points of K-Means,and use K-Means to automatically classify the image data set;then perform self-checking and beam calibration for each type of image Poor;multi-group camera parameters obtained by self-calibration are weighted average to obtain initial camera parameters;finally,global self-check beam adjustment is performed according to the initial camera parameters.Multiple experiments have shown that the pixel distortion of the self-checking and indoor calibration parameters is 0.5 pixels,the root-mean-square error of the checkpoint is 10.1cm,and it is better than Smart3D,Visual SFM and COLMAP software.It has the original pose that can more accurately represent the data.The classification self-checking and calibration method can provide an effective solution for the self-checking and calibration of UAV images in the strip area,and has strong practical application value.