Research On 3D Reconstruction Of Point Cloud From Terrestrial Laser Scanning

Terrestrial laser scanning technology, to which lots of improvements and breaking-throughs have been contributed, has gradually won more and more cases when applied to object reconstruction, and raises as as one of the fundemental data of the reconstruction works. By recording data like amplitudes of return waves, and colors of fusioned CCD images, this new terrestrial laser scanning technology retrieves information of the radiation properties of the surfaces of the objects, while traditionally only geospatial attributions could be distrilled from the array-like point clouds with 3D position coordinates. The geospatial data are always used to construct the neighboring relationships of points and topological natures of surfaces of the targets, based on which surface geometrical modeling could be built either as grids or as parametric curvatures. Meanwhile the radiation information are employed to recover the color and texture features of the models. Ever since the first introduction of augmented algorithms with auxillary threshholds provided by extra radiation conditions, the synthetic process has attracted quite a lot research effort and concerns, in the hope of achieving better robustness. After a well-structured review on classical theories and algorithms of key problems of terrestrial laser scanning, including matching, segmentation, outline extracting and texture mapping of the productions of the point cloud data, this paper, utilizing radiation information, makes several tries to improve ability of the algorithms, like adapting to larger application fields and stay sensitive even when dealing with tough data sets. Beside detailed discussion about system structure, scanning process, building neighboring relationship and topological nature, abd radiation data of the terrestial laser system, this paper concentrates on the following topics:(1) In a background that trigonal grids have been broadly accepted in everyday work, and aiming at difficulties in establishing homogeneous feature relationship, a new grid-feature based, auto-registration algorithm is proposed. Several remarks on this algorithm are:1. It filters featured point sets using neighboring average metric curvature method, which comes from properties of vertex curvature of trigonal grids, and resulting in more reliable curvatures; also, it forces the initial orientation of the clouds to the same direction interactively, and receives reduction of computational complexity of pre-registration.2. It engineeringly proves that applying ICP over pre-registration data sets, which are generated from vertex of selected trigonal patches computed from k-d tree, would contribute to a better registration.3. In order to avoid odd situations that curvature computation would fail occasionally, it sets up a filtering model according to radiation attributions of grid vertex, and eventually serves as a backup stradegy.(2) Starting with analysis of present segmentation algorithms, this paper focuses on the feasibility of hypothesis that additional conditions, namely amplitudes of return waves, would work with segmentation methods and enhence the output. In a knowledge that flat planes are popular in building facades and there is a apparent relationship between object reflection attributions and amplitudes of return waves, an improved plane-growing algorithm, which refines growing threshhold conditions also in consideration of amplitudes of return waves, is designed and furthermore, RANSAC is used to get better robustness on parametric estimation of initial patch models. Moreover, the corresponding tests prove that to point clouds under some specific conditions, modified method would cause oversegmentation and leaves with residue patches, however, it does contribute to improvements of recognition of object features as well.(3) After analyzing the classical structural and geometrical features of building facade’s point clouds, this paper discusses improved scanning strategy in fast rebuilding of building facades. Then, it focuses on methods that transform point clouds to outline models, and proposes job procedures on extract inner and outer outlines from nonuniformly distributed point clouds data, with combined methods of both border trace and spatial grids, due to effects on different methods caused by unbalanced data. This new method determines outline seed by local maximum and traces the border at a specific level; also, with the help of spatial grids, it detects the vacuums and chooses initial point set of inner outline. Because of inner outline cannot sincerely reflect features of facades, for example, the real size of window, this paper first fits the segments of the border point sets, and then correct the borders with measure data. Finally, it normalizes the border to get the outlines.(4) From coherence of scene lighting, this paper details on the necessarity of image dodging and how to choose dodging models. In this section, a series of overlapped images are dodged with Wallis filter, and coherence is archived on every color band of a pair of images. Accompanying with distortion models, geometric correction models, relative positions of camera, laser scanner and objects, two different projection models of the angle between camera and laser scanner are talked about. To the small angle situation, DLT method is accepted to transform point clouds to texture images, and inner orientation elements are solved directly with point clouds’coordinates. To the large angle situation, camera calibration is demanded and projection should be coped with strict model of resection method. In an conclusion, texture projections of above two models have been applied to several test data, and archived an improved result.The data processing method by integration of geometric topology information and spectral radiation data has great potential applications in the reconstruction work based on terrestrial laser point cloud data. Through studying several key issues on object 3D reconstruction, achieve certain results by analyzing the methods about using spectral radiation model to improve the quality of reconstruction quality and give a summary of the full text at last.