3D Reconstruction Based On Dense Matching Of Image Sequence

The technology of3D reconstruction based on monocular image sequence iscomplicated and the reconstruction accuracy is not high, however, due to its flexibleoperation, low cost, ease of installation and so on, it can be widely used for computervision, such as navigation, virtual reality, obstacle recognition, industrial inspection andgeographic measurements, which has extensive application value. This paper does theresearch on how to improve the accuracy of reconstruction.The main work is as follows:Access to a large number of literatures about reconstruction based on imagesequence home and abroad and discussed the research situation. Introduced the relatedtheory of3D reconstruction based on image sequence of dense matching, including thecoordinates, camera model, epipolar geometry, fundamental matrix, dense disparity mapand depth chart, etc.Based on the studying of the method of feature point extraction and matching,proposed a method of extracting features which based on pixel interpolation.Themethod in the case of the other conditions unchanged, Through pixel interpolation, theoriginal part of the seed point of sparse or no seeds in the region to obtain the seed point.After analysis the pros and cons of Sift matching method, introduced NCC matchingmethod to the Sift. This will not only accelerate the matching speed and it can also get abetter robustness, more uniform distribution seeds.To improve the reconstruction precision, design and implement a dense matchingmethod based on region growing. Solved the problem of lacking features, reconstructionlocalization, and even distortion and other problem. The method exploits the method ofextracting features, find the robust seeds in the image first; then take the seeds asstarting point, through a certain growth method to spread the matching relations to thewhole image; finally, using RANSAC role in the matches, which to obtain the finalmatches. Experimental results show that the matches obtained both in quantity anduniformity of the distribution are all better, the method improved the accuracy of3Dreconstruction.Design of3D reconstruction system. This paper designs the whole system workflow, and discussed many problems, such as the estimated motion parameters andoptimization,3D points computing, triangulation and the principle of texture mapping. Finally, developed a3D reconstruction software, besides, through experiments to checkthe feasibility of this method.