Three-Dimensional Structure Reconstruction From Aerial Images Based On Affine Approximation

In aerospace, industry inspection and disaster surveillance fields, we can use camera to capture images of targets or scene, from which three-dimensional (3D) information can be got through 3D structure reconstruction.3D image analysis is an inverse problem from two-dimensional (2D) image to 3D data, which always be ambiguity and sensitive to noise and digitization error since the projection model is complexity. In a long term, most learners use perspective model to describle camera projection model. It is necessary to close-range image analysis, but may lead unstability in aerial images (or far-distance images) analysis, as there are few perspective distortion.In order to 3D positioning for ground of unmanned aerial vehicle, the article presented a new method based on affine approximation. That is to say, initial value was often solved linearly based on affine model, followed by optimization based on perspective model. The article is composed of three parts.In the first part, affine projection model was studied, which gave a detailed study on the affine model, including its algebra and geometry form, and particular discussion on 2 typical affine models, weak perspective and paraperspective. Research on the error of affine model relative to perspective model shows it is reasonable to analysis aerial images based on affine approximation.In the second part, the camera calibration was investigated where control points are often far away from camera. A new calibration algorithm based on affine projection matrix, named APMC, was presented, as long as the view of being unable to use re-projection error to judge camera calibration algorithm lonely.In the third part, 3D structure reconstruction from aerial images, including sequential images and two images, based on affine approximation was introduced in detail. This is the primary of the article.On behalf of 3D structure from image sequence, a new algorithm based on affine factorization and bundle adjustment, named AFBA, was put forward. Being the sub-algorithm of AFBA, two new bundle adjustment algorithms, bundle adjustment without using camera extrinsic parameters and an improved quasi-linear bundle adjustment were worked out. If some image points are unseen, the measurement matrix comprises of some missing elements, which leads impossible to directly factorization in AFBA. In order to solve the problem, a new algorithm named ATIS was presented based on tri-linear tensor.On behalf of the 3D structure from two images, a new algorithm named AERO was putforward based on afirne epipolar constraints. In succession, a new method based on the restriction among essential entries was presented to solve those problems without known of the focal lengths of cameras.The four new algorithms, APMC, AFBA, ATIS and AERO are all based on affine approximation, which effectively solve camera calibration, 3D structure reconstruction from image sequence and 3D structure reconstruction from two images, respecitively. A great deal of real image and synthetic image experiments show they are better than traditional algorithms.Photogrammetry is consistent to computer vision in purpose and mathematic framework. The article gave a tentative research on the intercourse between them. Some new algorithms were realized by combining theories and algorithms both in photogrammetry and computer vision.