Research Of 3d Reconstruction Based On Computer Vision

In these years, more and more attention is being paid to the application of Computer Vision. 3D reconstruction from images is fundamental and has been extensively studied in Computer Vision. It uses multiple images to reconstruct a 3D scene. There is more and more demand for 3D models in robot navigation, visual inspection, computer graphics, virtual reality and communication. There is an important demand for simple and flexible acquisition procedures.This thesis investigates the reconstruction using multiple images of a rigid 3D scene taken at different viewpoints. We tackle and contribute to several problems in different stages of 3D reconstruction:1. Camera calibration. We introduce the definition of camera model and geometry in detail. In the paper, an optimization method based on Zhang's calibration is presented. It is used to undistort the image. The modification preserves many advantages possessed by the original Zhang's calibration technique.2. Feature extracting and matching. Firstly, we introduce the concept of feature and scale. Secondly, we give a brief overview of some of the important feature detectors and descriptors. Thirdly, to speed up the feature matching process and overcome the problem of the high dimensionality of the SIFT descriptor, an optimization method is presented.3. 3D reconstruction based on multiple images. Firstly, we introduce the theory for two view geometry, including the concept of fundamental matrix and essential matrix. Secondly, the RANSAC algorithm is used to cope with two view geometry computation. Thirdly, 3D reconstruction based on multiple images is presented in the paper. At first two images are selected and an initial reconstruction frame is set-up. Then we discuss how to add a view to the existing reconstruction and each time the initial reconstruction is refined and extended by bundle adjustment.In this paper a complete system to build 3D models from camera images is presented. The system can deal with image sequences acquired with a hand-held camera. It is based on the structure and motion recovery.