Research On Segment-based Dense Matching Algorithm And Its Realization In UAV Images

Unmanned Aerial Vehicle （UAV） low-altitude photogrammetry system, as a kind ofconvenient and low-cost data acquisition method, has shown unique advantages in thereal-time response and3D reconstruction, and it has been widely used. Reconstructing theobject’s3D information with UAV images is characterized by high automaticity andinexpensive cost. Dense matching technology is one of the key techniques for reconstructing3D information with2D image and also a hotspot and difficulty in photogrammetry andcomputer vision. Compared with traditional aerial images, UAV images have the advantagesof higher resolution and larger degree of overlap, and there are also many disadvantages suchas a smaller base-height ratio and unstable images gesture. These characteristics of UAVimages bring difficulties to the dense matching of UAV images.This thesis mainly research the dense matching algorithm of UAV images. By studyingthe classic dense matching algorithms at home and abroad, we analyzed and compared theadvantages, disadvantages, and the scope of application of the commonly used densematching algorithms. According to the characteristics of UAV images, we designed asegment-based dense matching algorithm for UAV images. The mainly processing flow is asfollow: firstly, a mean shift color segment method is used to segment the left epipolar imageinto regions with homogeneous color. secondly, a semi-global matching method is used todetermine the initial disparities of reliable pixels. Thirdly, a ransac plane fitting technique isapplied to obtain a set of disparity planes that are considered as a plane set. Fourthly, anoptimal disparity plane assignment is approximated using belief propagation optimization.Finally, a disparity refinement is applied to improve the disparity map. Considering theefficiency and adaptability of the algorithm, we used the strategy of block in this paper. In order to evaluate the accuracy of the method, we generated dense3D color point cloud toevaluate the precision of the algorithm in this paper, except using the disparity map.Experiments prove that the algorithm we proposed has good effect for UAV images and hasmake a big progress compared with SGM method.The main research contents and innovation points in this paper are listed as follows:1) We study the classic dense matching algorithms at home and abroad deeply, andcompare the advantages, disadvantages, and the scope of application of the commonly useddense matching algorithms. Combined with the characteristics of UAV images, summarizethe problems in dense matching of UAV images.2) This thesis has designed and realized a dense matching algorithm for unmanned aerialvehicle multiple images based on segment. The semi-global matching method is used thatreplace the local algorithm to calculate initial disparity map and improve the accuracy ofinitial disparity map. In addition, a disparity refinement was used at the end of the densematching process and improved effect of the disparity map, especially in the case ofinaccurate object boundary caused by image segmentation.3) In dense matching process we have used the strategy of image block. It solved theproblem of insufficient memory due to the excessive disparity space image and improved theapplicability of the algorithm we proposed.