Image-based 3-D Reconstruction Of Tree-like Objects

Virtual reality technology needs to acquire the 3-D model of the objects in a virtual environment. And image-based modeling is a method of modeling which has been attention-getting in recent years. It can get the realistic model of a specific object. The 3-D reconstruction of tree-like objects has important application in medicine, biology and the simulation of natural scene. This thesis expatiates the problem of 3-D reconstruction of tree-like objects from images and discusses the method of reconstructing the branch models of a tree from two photographs at different viewpoints. In the meantime, we probe into the problem of image segmentation and recognition related to the reconstruction, and discuss the problem of tracking of players from soccer video sequences. The main content of this thesis is as follows:1. The thesis presents a scheme of reconstructing 3-D model of tree-like objects without leaves from two photographs at different viewpoints. The scheme uses the technology of 3-D points from stereo. It reconstructs the 3-D skeleton from 2-D skeleton, and then adds the breadth information to get the 3-D model of the tree. It presents a method of automatically acquiring 2-D skeleton data of main branches. It also solves the correspondence of the 2-D skeleton points. And it realizes the reconstruction and simplification of 3-D points, and the generation of surface meshes of branches. The scheme preserves the quality of reconstruction with a little interaction. It is a practical and effective method.2. The thesis presents a method to implement quick segmentation and tracking of players from a soccer video. This method discusses the segmentation, team identification and tracking of players. (1) Players are segmented automatically by using the statistic of color components differences; (2) Correlation of two normalized color histograms is used to identify the team to which a player belongs; (3) To track the players, it has used the context features and correlated-template matching. The method realizes the real-time tracking of players successfully. It deals well with the occlusion of two players belonging to different teams, and can track when multiple players are in occlusion.