Study Of 3D Reconstruction Technology Based On Binocular Vision

Three-dimensional(3D)reconstruction from images is one of the major research in computer vision.We study the method of 3D reconstruction from two aspects: 3D reconstruction based on disparity maps and 3D reconstruction based on feature points.Then we improve some steps in the approach of 3D reconstruction,experiments and analysis related to these improvements is performed.With respect to 3D reconstruction method based on disparity maps,we propose a stereo correspondence algorithm based on guided filter and semi-global method.First,we improve the similarity measure function based on the pixel gray value and pixel gradient by combining it with a function which has strong anti-noise ability.Then improved similarity measure function is filtered by guided filter.We define an energy function by means of filtered similarity measure function.In order to get final disparity maps,we minimize the energy function mentioned above using Dynamic Programming method described in semi-global algorithm.The experimental results show that the stereo correspondence algorithm based on guided filter and semi-global method performs well on the test data set,but its effect on images captured by us is not satisfying because edges of objects in disparity maps is not smooth enough.Therefore,the generalization ability of the algorithm is not good.For 3D reconstruction based on feature points,we adopt SFM(structure from motion)algorithm for reconstruction.With respect to the circumstance that error matches in feature matching involved in SFM,we propose an algorithm for removing false matches based on pixel gray value,pixel gradient and pixel distribution in neighborhoods.First,the zero mean normalized cross correlation of the two feature points is calculated,and then the similarity value of the pixel arrangement in the neighborhood of the two feature points is calculated,finally the similarity value of the pixel gradient for two feature points is calculated.According to these three criterion,we can judge whether the two feature points are correct correspondence.The experimental results show that,compared with RANSAC,the number of correct matching pairs obtained by our algorithm is less.However,the running time of our algorithm is less than RANSAC algorithm.So our algorithm is more suitable for real-time application.