Study On The Key Technologies Of Building 3d Reconstruction Based On Multiple Images

With the rapid development of science and technology, civil and war industries demand three-dimension (3D) reconstruction more and more. To meet these demands, the technology and theory for 3D reconstruction have increased dramatically in the past few years. And that the 3D reconstruction technology based on images is a rising reconstruction way based on optics, electronics and computer technology, it fits to the modern production which needs reconstruction technology that holds flexibility and application from a bran-new idea.The subject is"Study on the key technologies of building 3D reconstruction based on multiple images". Multiple images of an object are captured by a hand-hold camera and used for 3D reconstruction. We need solve some key problems, such as edge detection, point matching, projective reconstruction and Euclidean reconstruction, etc.Edge detection is the precondition of 3D Euclidean reconstruction. The paper proposes an edge detection algorithm based on dyadic wavelet transform and morphological opening. Combined with dyadic wavelet transform and image gradient calculation, the four smoothing filter operators are introduced. Then a local thresholding, which determines the edges, can be constructed by the algorithm of maximization of between-class variance and morphological opening. The experimental results show the algorithm detects the image edge correctly and has the good suppressing noise ability.Point matching is the basis of 3D Euclidean reconstruction. The paper proposes a point matching algorithm for image pairs with noise based on the two-level strategy. Firstly, the modified sum of absolute differences (SAD) is used to measure the edge similarity within a large template window to obtain a coarse-level matching of points based on the contours of the images. Then gray similarities of the points are measured within a small region in the fine-level matching based on the result of the coarse-level matching to obtain the correspondences of the points. Combining the combination definition and symmetrical probability density function of noise, the order statistic filter (OSF) is improved to estimate the image intensity. The algorithm theoretically explains its simplicity and experimentally shows that the matching is not almost influenced by image noise under different signal to noise ratio (SNR) levels by experiment.Projective reconstruction is the necessary step of 3D Euclidean reconstruction. The paper presents two algorithms. One is projective reconstruction with missing data based on factorization algorithm. It estimates projective shape, projection matrices, projective depths and missing data iteratively. Estimation problems of projective shape and projection matrices are solved in terms of singular value decomposition. According to the fact that the sum of linear subspaces, each of which is spanned by the points of each image, is equal to the linear subspace spanned by space points, projective depths are computed. The other is projective reconstruction with missing data combining 2D reprojection error and subspace method. Estimation problems of projective shape and projection matrices are formulated in terms of the minimization of 2D reprojection error. The subspace method is used to estimate projective depths. The missing data can be updated by computered projection matrices, projective shape and projective depths. Experimental results with both synthetic data and real images show that the proposed methods have small reprojection errors, good convergence property and can be used for the condition that object points are missing on some of the images.At the transformation from projective reconstruction to Euclidean reconstruction, the paper gives an algorithm to reconstruct 3D Euclidean shape of the object from the projective reconstruction. In the real condition, we set that the focal length and the constant principal point are unkown. Using the fact that singular value decomposition of the scaled measurement matrix recovers multiplicate motion and shape pairs, the equation sets of the transformation matrix, which is inserted between projective motion and shape to get Euclidean motion and shape, are set up by enforcing the metric constraints. The algorithm is applied to real images captured by a camera. Experiment results show that the algorithm can give the good shapes of the reconstructed objects. And to further verify the feasibility of the 3D reconstruction, the paper tests the Euclidean reconstructin results of the real objects to obtain the length information of the objects by a scale factor, which is computed using Euclidean reconstruction results and real objects, and analyses the factors that cause errors of the reconstruction. Experiments prove that reconstruction principle is correct, the algorithm is reliable, the implementment is easy, the reconstruction can give the correct shape of the three dimension object in the case that object points are missing on some of the images, and is practically applicable.