Study On 3D Reconstruction Key Technology Of Unordered Multi-view Images

Reconstructing the three-dimensional structure of a scene from multiple two-dimensional images is a hot research topic in photogrammetry and computer vision. With the development of image capturing and 3D reconstruction technology, the effectual source image data that can be used for reconstruction is constantly expanding, which makes this topic has a broad prospect of application. This thesis is focused on automatic 3D reconstruction problem without any prior information of scene structure based on unordered images, reseaches and explores the related technology in the process of reconstruction aiming at improving the accuracy and robustness of 3D reconstruction. This thesis makes the following contributions:1. The construction method of epipolar geometry relationship between two images in the calibrated and uncalibrated condition has been studied systematically. The general equation of dangerous surface in the critical configuration is deduced using the coplanar condition function in the object space form, the critical configurations of the straight line and plane are analyzed, and the critical conditions of different construction algorithms is computed using singular value decomposition.2. Based on the analysis of feature detection and matching algorithm between images, information criteria algorithm is introduced to detect fitting model of matching data. The error equations of homography matrix model and fundamental matrix model in information criteria algorithm are derived, and the expectation maximization algorithm is used to deal with mixed model problem in calculation the standard deviation in order to avoid the possible degradation in the subsequent reconstruction processing, and to improve the reliability of 3D reconstruction result. The extraction problem of multi-view matching points is converted into the dynamic connectivity problem; a Union-Find method is designed and then a weighted coding strategy is used to optimize the problem. By these two strategies, the efficiency and reliability of matching point extraction are improved.3. A detection algorithm of false positive matching relations between images is proposed. The detection problem of false positive matching relations is converted into the probabilistic inference problem, probabilistic graphical model based on Bayesian network is constructed using closed cycle constraints in image relation graph and information obtained in the matching process. The image triplet closed cycles is added to the cycle constraints constructed by maximum spanning tree generated using greedy algorithm and the influence of the structure of image relational graph has greatly reduced to detection results, the false positive matching relations caused by texture repeat or similarity can be effectively avoided and the outcomes of scene reconstruction improved.4. A reconstruction algorithm of multi-view images using hierarchical clustering is proposed. The internal order of the reconstruction is optimized using agglomerative hierarchical clustering which can reduce the influence of initial image pairs to reconstruction results. The drift problem in incremental reconstruction method caused by error accumulation can be effectively avoided and the robustness of the algorithm of 3D reconstruction can be improved.5. A batch reconstruction algorithm of multi-view images based on image triplets is proposed. The image triplets which may contain mismatches are removed using closed cycle constraint, which can effectively avoid the false positive epipolar geometry relation, and then the trifocal tensor constraint in triplet is used to calculate a global optimum of rotations and locations of all the images to improve the reliability of the reconstruction results. Because of there is no need to have iterative adjustment computation, it has relative high computational efficiency.