Summary: | 碩士 === 國立成功大學 === 測量及空間資訊學系 === 102 === This study presents four methods for blunder detection and quality evaluation on dense matching results. They are visual check, relative orientation (RO) using a huge number of tie points, bundle block adjustment, and comparing with ground truth data. To detect blunders, not only 3D object points but also 2D image points need image information. Therefore, there are bottlenecks, which include (1) no original image coordinates of matching points, (2) a huge number of matching points, and (3) the matching points in close distance making the calculation of bundle block adjustment unstable. The most probable values of RO five unknown elements are calculated and sparse dense points are selected to solve the bottlenecks in blunder detection. In this study, test data are the results of four matching algorithms. They are SIFT-based Multi-image Matching (SMM), Structure from Motion (SfM), DAISY and Semi-Global Matching (SGM). The result shows that dense point clouds in areas with break lines, roof ridge lines and shadow are prone to have more blunders than other areas. According to RO computation, the matching error percentage is 2.82% and 2.36% by SMM and SGM, respectively. From aerial triangulation, the matching accuracy and error percentage of SMM are 0.23pixel and 3.97%. The computation speed of RO and bundle block adjustment are 14,984 points/second and 292 points/second, separately. We compare the dense points determined by SGM with ground truth data.The absolute elevation differences show that the maximum is 0.935GSD, minimum is 0.006GSD, average is 0.315GSD and root mean square difference is 0.238GSD, where GSD is 0.168m.
|