Robust method in photogrammetric reconstruction of geometric primitives in solid modeling.

Ph.D. Thesis Xiang Yang


The 3D point cloud is a widely used data format obtained from scanning a 3D model, either by using active 3D laser range scanners or passive photogrammetric methods. Since the topological information in a point cloud is captured on 3D point level, the inverse design cannot be carried out directly on the data. The point cloud is first segmented into various geometric primitives, such as planes, spheres and cylinders, then the modification and redesign of solid model can be more easily achieved.

A robust estimator is required to detect the multiple inlier structures while filtering out the outliers. In this dissertation, we present a new robust algorithm which processes each structure independently. The user gives only the number of elemental subsets for random sampling, which is also required in other robust algorithms. This method provides a general solution of robust estimation, and no tuning of other parameters are required for particular estimation tasks. The scales of the structures (tolerance of error) are estimated adaptively and no threshold is involved in spite of different objective functions. After classifying all the input data, the segmented structures are sorted by their strengths and the strongest inlier structures come out at the top. Like any robust estimators, this algorithm also has limitations which are described in detail.

To illustrate its efficiency and robustness, the algorithm is tested on various synthetic and real examples in both 2D and 3D. We extend its applications through the entire process of the structure from motion method, to reconstruct the 3D point cloud from a sequence of 2D images. We automatically estimate and fit the 3D surfaces from the 3D point samples, without generating surface normals or mesh model. The designer can interact with the 3D points conveniently and direct modification of point cloud becomes applicable.

The thesis contains 100 pages.

Return to Theses