Research On3D Scene Reconstruction Technique Based On Multiple View

As the developing of computer technologies, obtaining realistic3D scene information automatically plays more and more important role in our social production and life. Because multiple view contain more information and expand the scope of the field, so3D reconstruction based on multiple view can complete the work which a single view can not complete or complete difficultly.3D reconstruction based on multiple view is a hotspot research aspect in computer vision fields since it is inexpensive, convenient and simpler.There are many key techniques of3D reconstruction based on multiple view such as feature detection and matching, fundamental matrix estimation, camera self-calibration,3D scene reconstruction, dense surface reconstruction and so on. The thesis is focused on fundamental matrix estimation, camera self-calibration and3D scene reconstruction. The main research work are as follows:(1) Fundamental matrix estimationIn order to improve the accuracy of fundamental matrix estimation, an improved MLESAC algorithm is proposed. Firstly, according to the distances between the matching points and the corresponding epipolar lines, the superior correspondences are chosen, random sample consensus is adopted to sample the superior correspondences, and then the initial fundamental matrix is obtained using the Levenberg-Marquardt algorithm to minimize Sampson error; Secondly, according to the epipoplar geometry and adding constraints to detect matching points set, then the accuracy of matching points set is further improved; Finally, the accurate fundamental matrix is computed iteratively by using MLESAC. Experimental results show that the accuracy of our algorithm is improved, and the stability is good too.(2) Camera self-calibrationIn order to improve the accuracy of camera self-calibration. We propose a camera self-calibration method based on GA-PSO Algorithm. Firstly, the simplified Kruppa equations based on the SVD of the fundamental matrix is converted into the optimized cost function. Secondly, the minimum value of the optimized cost function is calculated by GA-PSO, then the intrinsic parameters of the camera are obtained. Because the GA-PSO algorithm combines with the advantages of genetic algorithm and particle swarm optimization, the accuracy of computation is improved. The experimental results show that the accuracy of camera self-calibration is improved.(3)3D scene reconstructionIn view of the problems such as low accuracy of traditional reconstruction method by minimizing L∞norm, being sensitive to outliers and low efficiency,3D scene reconstruction based on L, and space point classification(marked as" L1-SPC") is presented. Firstly, L1approach is used to remove the outliers. Because we don’t need to calculate the outliers in the following3D reconstruction, the accuracy of reconstruction can be improved and the computation time is reduced. Secondly, the scene points are divided into the two classes, the first class is the scene points which are visible in two-view; the second class is the scene points which are visible in more than two-view. The optimal triangulation method is used for the first class; The improved L∞-norm method is used for the second class. The gap between the upper and lower bound of bisection is kept as small as possible in the improved L∞-norm method, so the computation time is decreased. The experimental results show that the3D reconstruction based on L1-SPC method is higher accuracy and more efficient.(4)3D scene plane reconstructionIn view of the problems such as low accuracy of3D scene plane reconstruction by using tranditional3D reconstruction method,3D scene plane reconstruction of two new model based on multiple view are presented. One is the model of scene plane reconstruction based on minimizing the reverse projection error, the other is the model of scene plane reconstruction based on minimizing the transfer error. The first model considering the knowledge that the reverse projection lines not only intersect but also meet in a scene plane, so minimizing the reverse projection error in the scene plane. The second model uses the transfer relation between the image plane and the scene plane, so minimizing the transfer error in the scene plane. At last, the optimized value is computed by GA. The basic principle of two methods is same, and the difference between them is the computational complexity. The experimental results show that the accuracy of two methods is almost same and the accurancy of3D scene plane reconstruction is improved greatly.At the end of this dissertation, the main research is summarized. It makes out the main innovations and research achievements, and also points out the problems and issues which need to further research.