Research On 3D Reconstruction Tecnology Based On Stereo Vision

The stereo vision technique is a very active research focus in the area of computer vision, which combines theories and methods of image processing, computer vision, computational graphics as well as biological physiology and so on. It is through the analysis and processing of multiple images to obtain 3D geometry information of objects, and realistic achieve mimic the function of the human binocular vision. Stereo vision is to observe the same object in the scene from two different angles and to get 2D image in different perspective, and then use the imaging geometry theory to calculate the position deviation between image pixels(parallax), at last, it gains object's 3D information.Today, 3D reconstruction technology has been widely used in the industrial inspection, 3D measurement, virtual reality, chemistry, geology and physics and other fields, as can be seen, the research of the subject can solve many substantive issues. Therefore its more in-depth study will be a very important significance.Implementation of stereo vision technology needs several steps containing the image acquisition, image preprocessing, camera calibration, stereo matching and 3D reconstruction. The main research work of the dissertation is as follows:(1) First of all, we can get two images under different perspectives through the camera, then the image gray-scale transformation, median filtering and edge detection processing, make color images grayscale, and effectively remove the image noise, enhance image edge and details.(2) Camera calibration use variable learning rate and momentum vector algorithm to train BP neural network, avoiding the disadvantage of falling into local extreme, get internal and external camera parameters, and it does not require good initial conditions, not very complicated calculation processing, through training to meet the orthogonality of the rotation matrix. Its advantage has been verified by the simulation of synthetic data by adding noise, it has a strong practicality, robustness and high calibration accuracy. (3) Stereo matching is the most important and difficult step, first of all, it extracts the corresponding edge points in the left and right image by the Sobel operator, and then creates matching objective function, this paper use Artificial Fish School Algorithm to achieve optimization of the objective function, avoiding falling into local optimum, and to reach the global optimum, experimental results show that this method can achieve better stereo matching, thereby improve matching accuracy.(4) Finally, according to the results obtained by the above steps, to calculate the 3D coordinates of the object according to the basic principles of stereo matching, through the parallax obtained and using triangulation based approach to recover depth information to generate depth map in 3D environments.