Fundamental Matrix Calculation And Its Application For Stereo Disparity Estimation

The calculation of fundamental matrix is a very important issue in the field of computer vision. In order to remove the exceptional data and increase the estimation accuracy of fundamental matrix, this paper proposes a robust algorithm for estimating the fundamental matrix based on blocking. Firstly, a set of matching points is obtained by a classic corner detection algorithm and a cross correlation algorithm. Secondly, exceptional points are eliminated by random sampling based on blocking from the set of matching points to form an inner point set. Finally, two different strategies are used to process the inner point set. One is the classic non-linear iterative algorithm, which processes feature point positional errors within a certain range. An iterative solution is carried out on the inner point set according to their deviation to adjust weighting coefficients, i.e., smaller weighting coefficient for greater deviation and larger weighting coefficient for smaller deviation. The other one is the robust expansion point by point algorithm, of which the eight point initial subset is expanded point by point. Each new point is checked according to the sum of epipolar line distance of all the inner points. If the sum is not increased , the new point is added into the subset; Otherwise, eliminated. In a word, the essence of the algorithm is looking for a subset whose error is relatively small as the optimal subset to estimate fundamental matrix by the improved eight point algorithm. It can be seen from the experiment results that either blocking robust algorithm can eliminate some mismatching points, reduce the positional errors, decrease the bad effect of the exceptional data and improve the estimation accuracy of fundamental matrix. The robustness of the second strategy is slightly better.Disparity estimation is one of key techniques in stereo coding system. This paper introduces an epipolar line constraint DT-based(Delaunay Triangulation) stereo image disparity estimation algorithm. The first step of this algorithm is to describe the right image or reference image by a DT mesh. Secondly, the disparity is estimated by calculating the corresponding points of triangles using the theory of epipolar line constraint. Finally, the disparity of the left image is compensated by means of affine transformation. It can be seen from the experiment results that the PSNR of predicting image is increased and the subjective quality has a certain improvement.