Robust Processing Of 3d Point Clouds

One of the common approaches of obtaining the geometry information of the objects in the world is generating realistic and vivid 3D models with the aid of various kinds of acquisition devices. This approach has been widely used in Computer Graphics, Computer Vision, Archaeology, Topography, etc. The raw output of acquisition devices is usually represented by Point Clouds, i.e. a set of 3D points in the same reference coordinate, which can describe the spacial distribution and surface properties of the object. Although techniques for surface reconstruction from point clouds as well as point clouds processing have been extensively studied and many advances have been made in the past years, many academies and researchers have been working on the development and improvement of the key techniques of robust point clouds processing, in the hope of improving the quality of the data and rendering the application as well as prevalence of the 3D acquisition techniques. There are two reasons for that, one of which is the increasing of the demand on acquisition techniques from different fields, and the other is the limitations exposed by existing point clouds processing techniques, especially the deficiency of robustness of these techniques when dealing with the defects contained in the data because of the limitations of acquisition devices, acquisition condition and acquisition procedures, such as noise, outliers, holes, missing of sharp features and uneven distribution. This thesis has studied the key techniques of the robust processing of point clouds, including feature preserving point clouds consolidation, robust normal estimation, feature lines extraction and global optimal approximation using simple shape primitives. In summary, this thesis makes the following contributions:(1) A novel feature preserving consolidation method for point clouds is proposed to preserve the sharp features, which are capable of capturing the surface properties, while eliminating the defects in point clouds, including noise, outliers and uneven distribution during the consolidation. The proposed method enables the existing Weighted Locally Optimal Projection (WLOP) based consolidation to preserve features by using weight function defined on normals. With the help of normal information, points with similar normals in the local neighborhood will be assigned larger weights, while points with large normal deviations are regarded as outliers and get smaller weights. Thus the new position is mainly defined by neighbors from the same side of the feature. At the same time, a normal mollification step is added to the new consolidation method to get accurate as well as feature preserving normals while the distribution of the points is improved. Experimental results on synthetic and real-world scanned data show that compared with the original WLOP, the proposed method can achieve denoisng, outlier removal and better distribution density as well as feature preserving on the consolidated point clouds.(2) A robust normal estimation method for point clouds is presented in order to deal with challenges from noise, outliers and sharp features. The idea of multiple structure detection in the field of robust statistics is applied in the proposed method. A random sampling approach is adopted to generate a set of candidate planes in the local neighborhood of each point. Based on these planes, a robust unbiased noise scale estimator is used to compute the local noise scale, then a Kernel Density Estimation (KDE) based objective function is defined to find the best tangent plane. At the same time, the procedure for tangent plane selection can also be used to detect outliers contained in the point clouds. Experimental results show that the proposed method is highly robust to noise, outliers and sharp features, thus can obtain smooth while feature preserving normals for the point clouds. Compared with existing methods, the proposed method is able to directly obtain feature preserving normals while avoiding tuning of complex parameters.(3) A RANSAC based method is proposed to detect feature lines which can describe the geometry features of point clouds which usually contaminated by noise, outliers and data missing. The proposed method is specially designed for the point clouds scanned from buildings or mechanical parts which usually contain planar parts. Starting from many planes detected in the point clouds, the proposed method projects all the points onto the planes which can approximate them well. The points on the boundaries of the projected regions are used as candidates for the final features lines, which are detected by the proposed RANSAC-based line detection method with global constraints. Experimental results show that the proposed method is highly robust to noise, outliers and data missing in the point clouds, and can extract accurate feature lines. Compared with existing methods based on local analysis, the proposed method is more robust, thus can be applied to more point clouds with low quality which are difficult to other methods.(4) A novel method for approximating point clouds with simple shape primitives is proposed to achieve accurate approximation while avoiding over-fitting, which can hinder the approximation from capturing the structures of point clouds. The proposed method combines variational surface approximation and model selection, and formulates the approximation as a global optimal problem on the partition of the surface of the given point cloud, the types of employed primitives as well as the number of the primitives by considering the cost of types and number of the primitives as well as the approximation error. Then an approximate solution for this combination optimization problem is given. A redundant set of candidate primitives is created, and then each candidate is iteratively refined using model selection and variational surface approximation until convergence. Finally a global model selection is used to determine employed primitives. Experimental results show that this method can generate accurate representation of the given surface that maximizes the approximation ability of the selected primitives, while being robust to noise. Compared with existing methods, the proposed one can automatically determine the optimal partition of the given point cloud as well as the type of primitive for each patch in the partition and achieve a trade-off between the complexity of approximation and the approximation quality while being independent on the specified number of primitives.