Research And Implementation Of 3D Rconstruction From Uncalibrated Digital Images

As the development of network and science technology,3D reconstruction technology is widely used in many fields such as medicine, the digital protection of the culture heritage and life in all fields. But currently it is difficult to obtain 3D information; the number of 3D models can not satisfy the needs of practical applications. How to quickly and efficiently reconstruct 3D models becomes an urgent problem to be solved. This paper researches on 3D reconstruction from uncalibrated digital images. The main contents of this paper are listed as follow:(1) Implement Zhang's flexible camera calibration algorithm. Experiments show that the precision of this algorithm is higher, and it has been applied to 3D reconstruction.(2) Propose a stereo matching algorithm for 3D Reconstruction based on uncalibrated digital images. First, improve the Harris corners algorithm which has corner location error and corner clustering phenomenon. Then construct the Sift feature descriptor around the feature point for initial match, and remove the obvious outlier points according to the principle of slope consistency. Finally, use bidirectional matching to find unmatching points by using fundamental matrix. Experiments show that this algorithm can reflect the structure of object or scenes and the match precision is higher.(3) Propose a robust fundamental matrix estimation algorithm. Matching error and corner location error are two factors which affect the accuracy of the fundamental matrix estimation. First, eliminate the effect of matching error to the fundamental matrix estimation by using LMeds algorithm; then eliminate the points of matching error according to the initial fundamental matrix constraints; finally, eliminate the effect of corner location error to the fundamental matrix estimation by using M-Estimators algorithm. Experiments show that the fundamental matrix has higher precision and better stability by using this algorithm.(4) Implement 3D reconstruction. Get 3D point coordinates by using structure from motion method, and 3D reconstruction realized from Delaunay triangulation and texture mapping. Experiments show that this algorithm can get 3D objects or scenes which can reach photo-level visual result.(5) Design and implement the system of Reconstruct3D. In this system, the structure of objects or scenes can be effectively reconstructed, and it also verifies the practicability of the proposed algorithms.