The Calibration And Point Cloud Processing In 3D Reconstruction

Vision is one of the most important ways accessing to the outside information from objective world. It is usually done through vision, touch, smell and other sensory organs to obtain information from the external environment. According to statistics, about 80% of the objective world information are obtained through the vision . Human vision, including not only the acquisition of information, but also the whole processing of transmission,treatment,storage and understanding. After the emergence of Signal processing theory and computer, People are trying to use the camera to obtain the environment images and converting into digital signal, complete the process of vision information by computer. In the way, formed an emerging discipline-Computer Vision.The reconstruction of 3-dimensional objects is one of the most important applications as computer vision, in recent years , more and more peoples get the attention, especially in virtual reality,reverse engineering and military. Stereo vision is an important branch of the methods of 3-dimensional objects reconstructions. stereo vision is a way that obtain 3-dimensional geometric object information from a number of images (usually 2) . In the computer vision system, we take advantage of two relatively fixed position cameras, from different angles simultaneously getting two images of the same object, by calculation the disparity of spatial points in two images we can obtain 3-dimensional coordinates.In the reconstruction process, there are some critical steps required to complete:the acquisition of images,camera calibration,feature extrac- tion,Stereo Matching,Three-dimensional reconstruction and the point cloud post-treatment. The article describes in detail each step of the 3-dimensional reconstruction, complete the acquisition of camera parameters, as well as access to the latter point cloud data to simplify,de-noising and triangulation .First of all we should carry out the camera calibration to get camera parameters, the relationships between 3-dimensional geometric position of some point in the space and correspondence point in the images is determined by geometric model, such geometric model parameters are camera parameters , the processing of solving these parameters is the camera calibration.According to the disadvantage of traditional calibrations for high-precision equipment as well as the the disadvantage of self-calibration method of low precision , lack of robustness. in the paper, we propose a new calibration method based on a planar template. the method is based on Meng'method[14]. Concrete steps are as follows:Printing a graphic template and pasting it in a plane, the template is constituted by a regular circle and a number of straight-lines that through the center of the circle;two cameras shooting from different angles , a number of template images are obtained(at least three required);using hough transform to separately detect the circles and straight-lines of two images obtained from two angles, and implement least-squares fitting to reduce the errors;solving the camera parameters from the points' coordinates in same images and the relationship of one common point between two different images .Calibration algorithm based on the template is difficult to determine as precisely as possible the physical coordinates and image coordinates of conics and straight-lines in the images. Therefor, the production of the template demands that the width of conic and straight-lines are as thin as possible based that they can be extracted clearly . Followed by the template as much as possible near the plane level. because we use the random hough transform method, each conic and straight-line fittings need repeat many times in order to lower the randomly generated errors. After the obtain of the equations of conics and straight-lines , according to the property of projective transformation:Linear and conic at the tangent point remain unchanged under projective change. In this way , we can get the image points of infinity from a number of tangent points in two template images, fitting a linear equation(called vanishing line) of infinity lines from these image points, through solving the intersection of conic and vanishing line, we can get two image points of circular points. According to the property of projective transformation:the image of absolute conic contains all information about intrinsic parameters. We can obtain at least three images of template in different projective transformations, this will help to create a equations, through the least squares can we get the intrinsic parameters matrix.For the solution of the extrinsic parameters, assuming that there are four tangent lines which are parallelling to each others. According to the coordinates of tangent points and conics, we can easily get the equation of tangent line. Here we also obtain four pairs of tangent lines of two template images from two direct, according this extrinsic parameters can be solved linearly[2].Image matching problem is that according one point in the left image how to find the correspondent point in the right image. Its mission is to calculate the disparity between two images. The disparity image represents the correspondence between two images. And by the principle of triangulation can obtain the depth of the scene. According to the different constraint manner, M. Z. Brown[32]divide into two:One is partial matching algorithm that process a small region around the pixel;Another is overall matching that constraint canning line or the entire image. This paper analyzes the two algorithms' respective advantages and disadvantages and their scope of application.Assuming that we have get the 3-dimensional points , access to the data processing stage. the most things for the large-scale point cloud are simplification and de-noising. Here we propose a simple way to deal with the point cloud. The general idea of the algorithm is as follows:First of all. according to certain threshold point cloud data is put into different pieces, making each small pieces' cloud data is easy to deal with;we use minimum quadric fitting to each small pieces(using Euclidean distance measure for convenience);estimated value of each point'curvature, this is to make the benefits of point cloud data maintain the non-uniformity. the result is that there are more points around high curvature region, less points around flat parts. Since the algorithm is used in least-squares fitting approach, The method also can suppress noise to some extent. At last, in order to ensure the effect of point cloud simplification, we completed the triangulation of 3D facial point cloud model. From the experimental results we can see that our simplification method can flexibly deal with large-scale scattered data points, and reach the expected requirements.