The Study Of Image Matching And The Generation Of Dense Points

Close-range photogrammetry is one of measures which get the three-dimensionalinformation.It has many advantages of high efficiency, low cost, convenient. In obtainingthree-dimensional information of urban buildings process, close-range photogrammetry hasthree steps: camera calibration, image matching, three-dimensional information calculation.Camera calibration is an essential process of three-dimensional reconstruction. OPENCVand camera calibration toolbox of MATLAB both are matural and practical camera calibrationtools now. Although they are similar in calibration principle and calculating parameters, therestill are quite differences in the practical application.Based on the given camera calibrationtheory, this paper uses the two methods for camera calibration, and gives detail steps andcalibration results.In the end, we conclude the factors that affect the calibration results and getthe proper condition of camera calibration.Aiming at the registration accuracy of image matching is susceptible to repeated featuresof urban images，a sift-based image registration method with anti-interfeence of repeatedfeatures is proposed.The algorithm chooses the keypoints with multi characteristic scales andorientations to bidirectional matching.RANSAC algorithm based on affine transformationmodel is adopted forfurther elimination of mismatch without buiding the descriptor.At last，we match the keypoints with single characteristic scale and orientation. The experimentalresults show that this algorithm can restrain the interference of repeated features, obtain ahigher accuracy of matching effectively and cost less time. The algorithm has realisticsignificance for urban image matching.Do diffusion point on the basis of obtaining the initial seed point. The traditional methodis not controllable in diffusion. In light of this, The algorithm of regional growth quasi-densematching based on grid is proposed. The algorithm to some extent inhibited the proliferation of points which are not controllable. This will generate a dense disparity map image of building. Then three-dimensional point cloud is generated by using relative orientation and forward intersection.