Feature Matching And Prediction Of Fundamental Matrix For Three-Dimensional Image Reconstruction

One of the most important researches of computer vision is the realization of three-dimensional reconstruction through two-dimensional images. The realization of three-dimensional reconstruction via binocular vision depends on two underlying jobs: feature matching technology and prediction technique based on fundamental matrix.This paper reviews the history of three-dimensional image reconstruction, feature matching and fundamental matrix estimation technology. Then this paper dives deeply into exploring the performance of various common used feature-point extraction algorithms, as well as their advantages and disadvantages. This paper compares traditional Harris and SIFT algorithm in various levels and, according to the uniqueness assumption, proposes improvements to traditional matching algorithms and prediction algorithm of fundamental matrix. With these preparations, feature points’ three-dimensional reconstruction is reached. The major content of this paper includes:(1) According to the uniqueness assumption and neighborhood information of feature points, this paper comes up with normalized cross-correlation (NCC) matching criterion and compares it with traditional algorithms.(2) This paper proposes a SIFT matching algorithm based on least weighted square (WSIFT). First of all, this paper proposes a self-adaptive threshold to adjust the number of the feature points, this method improves the efficiency of the original SIFT algorithm. Secondly, weighted least square method, on the basis of epipolar geometry, is introduced to finish the secondary screening of the traditional matching results, which successfully reduces the wrong matching rate.(3) According to evaluation criterion of feature extraction proposed by relevant scholars, this paper puts forward a new matching evaluation criterion, of which the influencing factors includes:the number of correct and wrong matching points, as well as the number of feature points located in the common part of two images.(4) In order to predict a more stable fundamental matrix, this paper combines the hearts of iterative algorithm and RANSAC algorithm to build a new way of prediction. Comparing with RANSAC algorithm, the algorithm proposed in this paper successfully reduces the cycle number of the experiments. With the help of weighted least square, it utilizes the weights of the reciprocal of square residuals for prediction. The result shows the improvement in the efficiency of calculation when using the algorithm put forward in this paper.The experiments of three-dimensional reconstruction in this paper are based on 3-parameter model. And the paper utilizes MATLAB 7.1 to carry out the affine reconstruction of the simple geometric figures and real-life scenes. Experiments reveal that compared with the traditional algorithms, the algorithm adopted in this paper has proved to be feasible especially in improving the reconstruction accuracy and stability of three-dimensional reconstruction. Comparing with traditional algorithm, the algorithm in this paper also has successfully improved the reconstruction accuracy and stability of the coordinates of 3-D objects.