Study Of Post-processing Techniques For 3D Scanned Point-cloud

Recently, with increasing requirement of acquiring 3D complex digital objects from real world in computer graphics, virtual reality, computer vision, reverse engineering and so on, 3D scanning technology have been developed rapidly. But for the reasons of measurement errors, orientation errors and viewport self-occlusion, we have to solve some difficult problems such as noise removing, hole filling and overlapped scanned patches registration for the scanned point cloud.In this thesis, some significant post-processing technologies for raw scanned data acquired by 3D handheld laser scanner are discussed. Then we focus on the topics of global consistent normal reconstruction from unorganized 3D point clouds and registration between 3D scanned point patches.In Chapter 2, we propose a robust normal reconstruction algorithm to recover the normal information from the noisy point cloud. We firstly remove the outliers by efficient hierarchical clustering. Then a normal reconstruction algorithm based on the hierarchical clustering is developed according to the statistic analysis of the noise. Furthermore, we give an evaluation criterion of normal reconstruction to adaptively adjust the parameter of hierarchical clustering. In experience, our algorithm generates high quality normal information efficiently.In Chapter 3, we present a novel registration algorithm for 3D point cloud patches based on CSRBF. Firstly, we generate initial correspondences between the points in different patches. Then the correspondences are optimized by a physical based method. Lastly, we propose a non-rigid registration method based on CSRBF warping. The computation of warping can be accelerated by multi-level hierarchical framework. The algorithm can be further accelerated by hierarchical clustering.We conclude the thesis and discuss some research directions for our further works in the last chapter.