Research Of Self Calibration Technique Of Binocular Vision

With the development of industry and computer science technology, the application of computer vision in the industry area has been more and more significant. Binocular vision technology has been the most frequently used and well developed technology in the computer vision area. The binocular vision technology can reconstruct a scene by two photographs of the scene taking from slightly different view. To realize the reconstruction work the parameters of cameras and the projective matrix of the system have to be calculated and this work is called camera calibration. Currently the main technologies of camera calibration are traditional calibration and self-calibration, In traditional calibration methods calibration objects are required to be put before the camera, and the camera is calibrated based on the given information of the calibration object. The computing work in traditional calibration technology is simple and easy, but limited by the calibration object. The self-calibration technology needs no calibration object and its work is based on projective geometry and the constraints of epipolar geometry. The self-calibration is easier to execute and more adaptive for practical use.Two self-calibration methods based on projective geometry and statistic theory are proposed in this paper. The first self-calibration method is inspired by monocular vision and is mainly based on projective geometry. When there are at least three orthogonal vectors in the scene, this method can calibrate the camera by analyzing the information contained in corresponding vanishing points and three selected matching pairs of points of two images. The second calibration method in this paper is called weak calibration. The method proposed in this paper improves the accuracy of fundamental matrix which can be used for reconstruction work by setting reasonable constraints. The improvement is based on the probabilistic model proposed by Mosian. In the new probabilistic model proposed in this paper, gray scale matching is introduced besides the F-rigidity of a set of points in order to find the most meaningful matching set to estimate the best fundamental matrix F. This method can work well in the case when there are some characteristic points are detected on two images but the matching relationship between point and point is still unknown. When there are a large proportion of outliers in the points detected, the method can still work well by setting reasonable gray scale constraint and F-rigidity according to the probabilistic model and find the most meaningful matching set and F.