Study On Quality Evaluation And 3D Rapid Reconstruction For Terrestrial Laser Scanning Point Cloud

As the rapid development of laser technique,terrestrial laser scanning(TLS)has been one of the most important methods to acquire 3D geospatial data which results in increasing wide applications in various fields.Recently,evaluating the quality of TLS point cloud has been rarely conducted that led to shallow researches on the precision of point cloud.Compared with the traditional methods to obtain 3D geospatial data,the advantage of TLS is the easy acquisition of dense point cloud that would be helpful to describe the surface of the scanned object.However,the disadvantage is the difficult determination of point cloud quality based on the point precision evaluation.Because the true location on the scanned object of a point in point cloud cannot be determined,the precision of a point in point cloud could not be used to demonstrate the quality of the whole point cloud.Hence,it still needs further researches to evaluate the quality of point cloud.In addition,3D reconstructed model of the scanned object should be obtained"finally after the data processing.Nowadays,the requirement of 3D model is increasing owing to the high-speed development of the society,it has been a hot issue to reconstruct 3D model rapidly from point cloud.Therefore,this paper focuses on quality evaluation and 3D rapid reconstruction for TLS point cloud,the details are as follows:(1)The sources of errors in TLS point cloud are firstly discussed systematically according to two classification methods.In order to eliminate the effect of systematic errors on the quality of point cloud,a new self-calibration method for TLS system based on automatic optimization of parameters is proposed then.The mathematical model of the systematic errors based on 21 additional parameters(APs)in TLS system is established,the procedures of the self-calibration is elaborated and the APs are chosen and optimized by determining a dimensionless quality index.At last,indoor calibration experiments have been carried out,the results shows that a group of APs with low correlation and high significance could be obtained with the proposed self-calibration approach,and the scanning precision could be improved highly after correcting the systematic errors using the obtained APs.(2).According to the theory of error ellipsoid,the procedures to estimate the error space of a laser point are stated considering the influence of both the laser measurements and the laser spot.Then the estimations of the actual point cloud error space and the actual point error space are derived by eliminating the overlap between neighboring point error spaces.Then a new method to determine the deformation monitorable indicator(DMI)is proposed by the relationship between the actual point cloud error space and the point error space.Finally,the performance of proposed approach is illustrated with validation experiments of planar board displacement where deformations with different sampling intervals,different scanning distances and different incidence angles were simulated.From the analysis of the experiments,the results show the validation of the feasibility of determining the DMI by the proposed method.This technique was also applied to a monitoring event of bridge pylon,and the results confirm the feasibility of the DMI in a real case,as well.(3)Due to that TLS point cloud just includes three-dimensional coordinates and reflected intensity,topological relations between points in point cloud should be firstly re-built,so the establishment and query of the local neighborhood are introduced and a kd-tree based method is emphasized.Then the methods to estimate normal vectors and curvature of point cloud are summarized and contrastively analyzed.The current methods of normal estimation are classified into three categories,whose strengths and weaknesses are enumerated according to the abilities to deal with the noises,outliers,sharp features and so on.The effect of a fixed neighborhood size on the results of normal estimation based on principal component analysis(PCA)is pointed out,and a new method using an adaptive optimal neighborhood is proposed.For a point in point cloud,a covariance matrix is firstly constructed by involving its neighbors,the establishment of the entropy function based on the dimensionality features comes next,the optimal neighborhood could be determined adaptively by minimizing the value of the entropy function,and the normal vector is finally estimated using the PCA method and the optimal neighborhood obtained just now.Two set of experiments have been conducted using simulated point clouds and scanned ones,respectively,and the results demonstrate that the proposed approach could improve the precision of normal estimation for the point cloud including sharp features.(4)To enhance the utilization of point cloud and reduce the redundancy in the data,simplification methods are discussed.Two common methods are firstly introduced and the disadvantages of them are pointed out.The method to extract features from point cloud is studied to preserve the important features in point cloud.Extracting the sharp points is carried out by setting up a curvature measurement formula.And the edge points are obtain by proposing a rapid method combining the k-means cluster and quadrant recognition for the point cloud containing clear boundaries.After the features extracted,a new progressive simplification method using the local entropy of normal angle for the non-feature point cloud is presented.The calculation of the local entropy of normal angle is derived,the procedures of progressive simplification are elaborated,as well.Then the local entropy of normal angle is also used to evaluating the performance of point cloud simplification methods.Three sets of point clouds have been used to implement simplification experiments with four methods in the end.The results of the proposed method show that more features could be preserved during the simplification,and the points appears to be uniformly distributed and no holes exists in the simplified point cloud.(5)The theory of parallel processing on GPU is introduced,and the preponderance of the GPU parallelism is highlighted in the beginning.The basis of Delaunay triangulation is discussed,and three methods to construct the Delaunay mesh are expounded,respectively,accompanied by the analysis of their merit and demerit.A new method of rapid triangulation for point cloud on GPU is put forward.To execute the proposed triangulation,an approach for parallel subdivision adaptively for point cloud is carried out firstly to obtain many subsets of the data,and a binary tree is then built for the subdivision.Parallel triangulation of the subsets is accomplished with the divide-and-conquer strategy,and the final whole mesh is obtained by merging the adjacent subset-meshes based on the binary tree.Then according to the respective superiority of the linear octree and the pointer octree,a GPU octree based on the level-order traversals is presented,and the steps to build the octree on GPU are elaborated.On the basis of the built octree,a GPU-based method of Possion reconstruction for point cloud is conducted,and important procedures on GPU are also given.In the end,experiments of reconstruction for seven sets of dense TLS point clouds that have different densities have been implemented.The proposed GPU-based methods could enhance the efficiency of 3D reconstruction of dozens of times and even one hundred times,compared with the traditional CPU-based method.