Research On Terrestrial LiDAR Point Cloud Data Registration And Fusion Method Of Point Cloud And Images

Many holes and imperfection of the object can exist in point cloud because localobject can be sheltered and it is difficult to set scan station at particular zone. Aimingat these questions, hole repairing of the uniform point cloud on account ofphotogrammetry is put forward after muti-view point cloud is registered precisely.The theory of point cloud registration and technical feature of photogrammetryare analysed deeply. Point cloud registration algorithm is improved to achieveautomatic and refined registration. Next fusion method of the point set from asequence of overlapping images and point cloud from laser scanning is exploredseriously. On the basis of laser point cloud preprocess，two registration algorithm isproposed. The first one takes advantage of geometric attribute implied in point cloud.Registration algorithm based on extracting and matching curvature feature is proposed.The other applies covariance eigenvalue to registration. Energy intensity of theneighboring point set is considered as registration element. Originally uniform pointcloud is achieved by means of point cloud registration algorithm above. Stereo imagescan be filmed around holes by analyzing empty point cloud. Then SfM theory can beapplied to several images to generate the relevant point set in photogrammetriccoordinate system. In order to achieve the integrated and uniform point cloud, thefusion of original point cloud from scanner and point set from photogrammetry seembe necessary. As a result of filling holes,3D point set is organized to model object inGeomagic software. Finally, feature is extracted based on Gaussian curvatureextremum algorithm. The major work and conclusions of this dissertation as follow:1. Introduce the process of acquiring point cloud and analyse the process ofpoint cloud preprocessing detailedly. Denoising algorithm is considered as animportant and unavoidable step because noise is bound to mix with point cloud. Thenstepwise least squares fitting algorithm is available for denoising terrestrial scanningpoint cloud. The feasibility and popularity of this improved algorithm is proved. Thenexplore and compare several algorithms calculating normal vector on points. Theneighborhood size of every point is the issue of great importance to the futurepoint-based data processing. So, the relation of neighborhood and curvature is takeninto consideration early. The relation curve is described.2. Two factors consisting of efficiency and precision based on intense researchinto classical registration algorithm need to be considered. Delaunay triangulation and curvature weighting effectively improve the algorithm. Next, RANSAC algorithm isconnected with coordinate transformation. The error matching point pair can beremoved so that the transformation parameter is more accuracy than originalalgorithm. The exsiting and improved algorithms above provide the foundation fortwo automatic registration algorithm: point cloud registration algorithm based oncurvature and the registration based on covariance eigenvalue. Experiment shows thatthese algorithms are efficient, stable and precise.3. Overlapping images coming from amateur camera can repair holes of pointcloud based on photogrammetry technique. So, using these images to generate3Dpoint sets is the key and necessary step. This problem can be solved by Structure fromMotion theory in this thesis.3D point sets in photogrammetric coordinate system canbe gained based on some existing program and improved algorithm. Point cloud holesand missing station result in incomplete point cloud.3D point sets coming fromimages are used for filling holes or registrater point cloud of empty scanning station.The fusion of two point sets can be achived by3D scale fator iteration algorithm. Theexperiment shows this algorithm is more efficient than scalar scale fator iterationalgorithm.4. Aiming at complexity and redundancy of point cloud data, point cloudcompression based on feature is studied. Firstly, candidate points are extracted byGauss curvature extremum. Secondly, reject the points in the approximate plane bycomparing adjacent normal vectors. Finally, several data sets are applied to thisalgorithm to verify the feasibility. This improved algorithm can extract point cloudfeature and compress the massive data effectively.