| Three dimensional point cloud model, using discrete point as a primitive, has a simple data structure and can express any complex details, it has been widely used in many areas such as reverse engineering, virtual reality, heritage protection. Due to the physical limitations of 3D laser scanning equipment and the environmental impacts of acquisition condition and procedures, point clouds often contain defects such as noise, uneven distribution, missing features. Many researches have concentrated on point cloud process in order to improve the acquired 3D data quality.The digital processes of complex buildings such as factory are often done by 3D laser scanning equipment. In order to collect the data of buildings as complete as possible, the buildings are not only to be scanned in plurality of different stations,but also the data of single acquisition is huge.Existing point cloud processing algorithms usually act on small scale point cloud models, which is difficult to directly apply to the large scale point cloud. This thesis has studied the key techniques of processing of large scale point clouds models, including point cloud simplification, automatic object extraction, point cloud resampling, curve skeleton extraction.In summary, this thesis makes the following contributions:(1) large scale point cloud simplificationA novel simplification method for large scale point cloud is proposed. Firstly,adaptive spatial division of point cloud model has been implemented by the size of division cube which is estimated by size of the bounding box of the whole point cloud, spatial distribution of points and number of neighborhoods. It has improved the speed of locating points, and nearest neighbors searching method is improved by using local spatial points density to optimization of spatial search direction; then the point cloud has been divided into different feature regions by the normal of points;Finally, region-dependent simplification was achieved by dynamical simplification parameters which were adjusted according to the feature information of the segmentations. Experimental results show that the proposed method avoids the defects of traditional algorithms, which pursuit simplified rate with serious loss of details or focus on preserving features with low simplified rate, the method not only has achieved balance of simplification rate and preserving features rate, but also has a fast simplification speed.(2) resampling method for point clouds based on feature preservingA novel resampling method for point clouds is proposed to preserve the features.The points have been classified into feature points and non-feature points by initial normal which was estimated by PCA. The normal of feature points have been re-estimating by iteratively local plane refitting with weighted normal and projected distance to the fitting plane which was formed by anisotropic neighborhood points.Then weighted locally optimal projection combined with weighted normal function was proposed, the project points were uniform distributed in the local neighborhood which have points with similar normal, and concentrated in the area which have large deviation of normal. At the same time, the new points were inserted in the sharp feature regions, the new points have evenly distributed in the local neighborhood and perpendicular to the underlying surface by the accurate normal.Experimental results show that the proposed method is not only to generate uniform distribution density as well as feature preserving on the point cloud, and it has effective hole-filling results.(3) robust curve skeleton extraction of point cloudA robust curve skeleton extraction algorithm was proposed. Firstly, the point cloud has been divided into "weak" strong convex patches based on computation of inner visibility between points in the models, the volumetric dissimilarity between the patches were calculated by earth mover’s distance using shape diameter function of points, the patches which had been classified into same semantic shape were fused and formed the collection of weak convex regions. Then the candidate skeleton points were extracted and compressed, smoothed and re-centered in each weak convex region. Finally, the branches of skeleton were generated by connected in each region, and the final curve skeleton was connected by the branches with the topology information of the segmentations. Experimental results show that compared with other algorithms, the proposed method can preserve the topology structure of original model with the connection information which was provided by the weak convex regions with lax semantic. It not only can robust extraction skeleton from the point cloud including different geometry, but also can robust extraction skeleton from incomplete point cloud with noise.(4)automatic object extraction from large scale point cloudThe method of automatic extraction of the classical primitives such as wall and cylinder from the large scale point cloud is presented. The seed points were randomly selected with normal weights and spatial density, the primitive fitting and detection were fused in the procedure. Then the light visibility tests based on OBB of vertical planes were computed, the fake walls were removed and potential wall were constructed by the results of the intersections of the visibility test. The over segmentation wall were fused by the normal of walls and distances. The detection of cylinders in 3D point clouds had been transformed into detecting circle in 2D space by respectively projecting point cloud onto different coordinate planes. And the remainder cylinders were detected by gauss map and iteratively fitting to get the better match parameters. The parts of the same cylinder were connected by mean shift clustering. Experimental results show that the proposed method can automatic extract the walls in the same floor and multistoried building. Compared with existing methods, the proposed method can extract as much as cylinders in large scale point cloud, especially the incomplete point cloud, but also robustly extract different types of cylinders. |
PDF Full Text Request