3D Model Reconstruction From Image Sequence

Three dimension model reconstructions from image sequence is a fundamental problem in computer vision and have been studied extensively in the last decades.Researchers are always trying to find a method which can build accurate models easily and quickly.However,there are many factors which can influence the efficiency and accuracy of the algorithms,such as various noises,occlusions,shape of the object and so on.Estimation of the fundamental matrix between two views is of great importance for many image sequence based reconstruction tasks.Due to noise and outliers contained in the set of putative correspondences,it is difficult to estimate the fundamental matrix with high accuracy and efficiency.To tackle this difficulty,by convex hull shrinking with angle constraint(CHSAC),we develop in this thesis a novel robust algorithm for estimating the fundamental matrix.With the shrinking of convex hull of correspondences in the image pairs,we obtain a subset of correspondences with much less proportion of outliers than the initial data set.Based on this subset,the putative matching points are weighted,an initial estimate of the fundamental matrix is generated from these weighted points,and a weight list of all correspondences is obtained.The final fundamental matrix is estimated in a robust way from the samples optimally selected with an angle-constraint technique from the list.Experiments on both synthetic data and real image pairs show that,compared with existing methods,our algorithm can enhance the precision of the estimate to the fundamental matrix with a reasonable amount of computation.The novel convex hull shrinking method with angel constraint technique(CHSAC)can be used in the 3D reconstruction pipeline to improve the accuracy and robustness of correlative algorithms.The method is more robust than the others and can get more accurate results which will be demonstrated in our experiments.CHSAC is embedded into the reconstruction algorithm to get the 3D model of the object.In the step of 3D model visualization,we generate triangular mesh for the models.3D oriented points set is need for generating the mesh as input for many algorithms.As the points set obtained by triangulation contain only the location,we need to compute the normal vector for each point in the space.In this thesis,we also proposed new method which is easy to understand and implement to compute the normal.