Automatic Creation Of Point Cloud Model For Terrestrial Laser Scanning

Light Detection And Ranging(LiDAR) provides an efficient and new solution for capturing geo-spatial information data with high space-time resolution in a non-contact and active manner. The ground LiDAR system which integrates multiple sensors, can obtain fine point cloud in stationary state. It has been widely used in the fields of quality detection, deformation monitoring, cultural heritage document and become a major means to obtain the geo-spatial data. On the one hand, it has a high speed in dense point cloud acquiring; On the other hand, the presence of noise and occlusion in the dense point cloud has brought huge challenge for point cloud data processing.The main problems in ground LiDAR data processing are the low efficiency from the large data size and high redundancy, the low automatic degree of point cloud registration for the artificial registration manner, and the poor generality in the existing automatic registration methods.For the problems mentioned above, this paper mainly put focuses on the method of automatic registration of point cloud and redundancy removing from multiple overlapping point cloud. The concrete research contents are as follows:(1) On the basis of introducing the fundamental theory and the data feature of the ground LiDAR, this dissertation presents two-level strategies to deal with the noise of isolated datasets to reduce its impact for the subsequent data processing, firstly, a distance threshold is set according to the approximate size of the object to filter out the noise in the object's surroundings. Then the occluding edges are detected according to the set threshold and spatial cluster is implemented for the edges in the raster structure, the isolated data areas are identified and filtered out according to the predefined size of the dataset.(2) Introducing the concept, method classification and state of the art of pairwise registration, this paper points out the drawback in the feature-based method and ICP method, stresses the importance of target-assisted method in practical engineering applications, and proposes a sphere target detection method for its isotropic, easy placement and high precision, which is based on occluding edges detection and multi-levels geometry constrain; Then, triangles constructed using three arbitrary non collinear sphere centers in each scan station are selected as registration primitives and the area and interior angles of each are selected as similarity measures. Finally, the congruent sphere centers between two scan stations are matched in an iterative manner and the transformation matrix is obtained by Rodrigues method.(3) The concept, method classification and state of the art of global registration is introduced. For the high memory consumption in simultaneous registration, this paper proposed an indirect adjustment method which is based on the independent model expressed by rodrigues matrix linearization and the initial value obtained in the pair-wise registration step, the final results are obtained.(4)The QMBB data structure is proposed to organize the single station point cloud and the3D virtual grid for the whole dataset. In each local grid unit, the corresponding sub-datasets are loaded dynamically, the redundancy is removed according to the distance to the scanner center and the point's normal vector.3D virtual grid converts the global redundancy problem to a local problem, which greatly saves the required huge size of memory.