A Multi-view Stereo Matching Method Of 3-D Reconstruction

With the development of the computer vision technology,the 3D reconstruction technique has been widely used in computer-based event simulation and visualization applications such as flight simulators,visual maps,games and so forth.Multi-view stereo matching method,as one of the 3D reconstruction methods,reconstructs the model in the scene using the information getting from a series of images.Recently years,the Multi-view stereo matching reconstruction method becomes a hot research subject in computer vision because of its convenience and feasibility.Aiming at some shortages of the Plane-Sweep algorithm,improve this algorithm basing on the original one.Combing the improved algorithm with the structure from motion method, a method of generating dense point cloud is achieved.Firstly we extract the SIFT points from input images.The method reconstructs the camera positions and gets the sparse point cloud of the scene simultaneously using a iterative method based on SFM which takes several images into account during each iteration.For optimization,we use bundle adjustment to get a better result.Subsequently,we get the information about the reconstruction model position,and with this position information we use a local plane-sweep method to get a quasi-dense point cloud of the scene.We choose SSD(Sum of Squared Difference) with an iterative sigma clipping technique during computing score stage of plane-sweep in order to deal with the scene occlusion.Then we filter most of the outliers of the quasi-dense point cloud by a combined mean and median rejection filter followed by a filter method based on NCC(Normalized Cross-Correlation) for the purpose to handle the variance of the illumination.Finally we merge all the quasi-dense point clouds computed from different views and get a dense point cloud.The final result of this reconstruction method shows that this method can deal with some occlusion and reduce the bad influence of the textureless background simultaneously and generate a dense point cloud with few outliers and high accuracy using multiple images of the reconstructed scene.