Facade reconstruction from terrestrial LiDAR point cloud is affected by both quality of point cloud itself and imperfectness of object recognition algorithms. In the facade, there are many features, which are shared within an object or among objects. For example, within one windo
...
Facade reconstruction from terrestrial LiDAR point cloud is affected by both quality of point cloud itself and imperfectness of object recognition algorithms. In the facade, there are many features, which are shared within an object or among objects. For example, within one window, the boundaries of it may share parallel and orthogonal orientation; windows in the façade may share same shape and orientation. This thesis focuses on these features, which can be called regularities, to improve the quality of facade reconstruction. There are three main objectives in the research. The first is to detect the wall and its holes, which represent existence of windows, intrusions or extrusions on the wall. Second, local regularities within one hole and global regularities shared among holes are explored. Last, the quality is evaluated by testing two datasets in order to show overall effect of the procedure. The RANSAC plane fitting algorithms and prior rules are used to find the wall planes from the point cloud. Instead of extracting holes directly from 3D points, a rasterization approach is employed to extract holes, a more robust way to deal with noise and various densities without loss of information. Regularities can be classified into two categories: local regularity with one hole and global regularity among holes. Global regularities consist of two types: global regularities among similar holes or between different holes. Nine cases of regularities within these categories are listed and explored. A hierarchical clustering is employed to identify regularities in certain feature spaces where the objects sharing the regularities will be close. We propose procedures to focus on finding features for corresponding regularities so that clustering can be employed to identify and apply them. In the beginning of the procedure, ICP (iterative control points) is used to identify groups of similar holes. The registered points and transformation matrices are used for finding features for identifying and applying local regularity and first type of global regularity. The boundaries from different holes are used to extract features in order to identify and apply second type of global regularity. To test the performance of the algorithms, two datasets from terrestrial LiDAR point clouds are used. The local regularities and first type of global regularities are explored for both datasets. The second type of global regularities is explored only for the second dataset, as there is only one group of holes in the first dataset. The quality of both results presented by holes on the wall is analyzed. The two results both show good effects of applying regularities and good matching with original point cloud.