Research On The Key Technologies Of 3D Reconstruction

3D reconstruction from image sequence from a uncalibration camera and some critical issues including projective reconstruction under both orthographic projection and projective projection, affine reconstruction, camera self-calibration and non-rigid reconstruction related are researched in the thesis. Most of traditional 3D recostrucion methods are based on a calibration image sequence, in which the camera parameters are kown.But in fact, the camea parameters are unkonwn in general, which limits their application. Without any prior knowledge either about the cameras, or the scene, how to reconstruction 3D object and the motion from an uncalibration image sequence is researched. Main contributions are as follows:1. To recover the 3D scence, a 3D reconstruction algorithm based on 1D subspace is presented under orthographic projection, which is based on the facts that the space points span the same linear subspace as the image points and that the two rows from the first image and a row vector which is orthogonal to the former can constitute a base. The row vector is obtained by iteration, and lastly the 3D reconstruction is accomplished.2. Ridding off the effect of the center point, we present two occlusion recovery methods based on 3D subspace and rank 3, respectively, under orthographic projection. The first method mainly relies on the facts that the rows in the matrix including all the image points span the same subspace as the rows in the matrix including space points and a linear iterative method is presented. The projective matrix that is from all the image points including all the occlusions that compose a matrix with rank3 is used to find all the occlusions in the second method. At the same time, the obtaining occlusions are substitute for the old occlusions. After some iteration, the real positions of the occlusions can be found. The innovations of both methods are that the positions of the occlusions are from all the visible points and that all the images and all the image points are treated uniformly.3. A 3D reconstruction method based on orthogonal complement subspace is presented. Two images are combined into a group and the occlusions in the group are removed. Then, the projective reconstruction is linearly obtained, based on the fact that the sum of orthogonal complement subspaces spanned by all the image groups equals to the one spanned by the structure. And, depended on the structure restriction, the affine reconstruction is upgraded. Lastly, the metric reconstruction is obtained based on the orthogonal projection model restriction. Being linearly, the method for 3D reconstruction overcomes the shortcoming that some existing iteration methods need good original value.4. Occlusion recovery algorithm based on organizational evolutionary is presented. The occlusions are regarded as the variants and the objective function is obtained based on the property that the rank of the matrix consisted of all the image points is 4. And the value searched by organizational evolutionary is taken as the occlusion position. The individual of population is evaluated by the objective function. And after several generations, the best individual is regarded as the occlusion position. The algorithm does not need the initial value and treats all the images and image points uniformly.5. A method for occlusion recovery based on rank 4 is presented. Firstly, all the occlusions and deep factors are assumed known to obtain a rank 4 matrix that is used to structure a projective matrix. Then, all the occlusions and deep factors are obtained from the projective matrix. At the same time, the obtaining occlusions and deep factors are substitute for the old ones. After several iterations, the real values of the occlusions and deep factors can be found. The innovation of the algorithm is that all the images and all the image points are treated uniformly.6. Based on the two pair parallel lines that are perpendicular to each other, a camera self-calibration method is presented in the paper. Depended on the facts that the model constraint can provide an equation and two equations can be obtained from two pair of parallel lines that are perpendicular each other, the self-calibration can be realized form only two images. The method presented in the paper is quasi-linear.7. To reconstruction the 3D non-rigid from an uncalibration image sequence, a 3D non-rigid reconstruction method which is based on factorization is presented. Firstly, based on the rank constraint, all the depth factors are obtained by linear iteration. Then, a reconstruction which is different a transformation matrix with the real one is obtained based on factorization. The transformation matrix can be linearly solved by using the constraints of the projective matrix, and the reconstruction can be upgraded to the Euclid one.