Point Cloud Registration And Extraction Of Typical Urban Features

With the maturation of three-dimensional laser scanning technology,three-dimensional laser point cloud data has become more and more widely used in urban planning,construction,and management due to its high accuracy and fast acquisition speed.However,the point cloud also has the disadvantages of high noise interference,limited single acquisition area,and complex identification algorithms,which brings certain difficulties to the research.How to solve the quick splicing of point cloud data and automatic extraction of geographic elements for the limitations of point cloud data is an important topic in the research of point cloud.This paper starts from the above two links,and on the basis of solving the problem of point cloud data registration integrating efficiency and precision,it fully exploits the potential point cloud information through the fine feature expression of point,and realizes the high-precision extraction of urban roads and street trees based on MLS data.This paper mainly completed the following work:(1)For the drawbacks of traditional Iterative Closest Point(ICP)algorithm in point cloud registration,ICP algorithm is verified based on the three important parameters of the coincidence degree,the include angle and the separation distance of the two point cloud datasets,and whether the algorithm falls into a local optimal value is used as the basis for judgment.Then,the changing rules of ICP under different parameter thresholds are summed up and the effective parameters range for which the algorithm can obtain correct registration results are given,which provides a reference for whether it is necessary to increase global registration before performming ICP.In addition,the accuracy and the efficiency are evaluated for point cloud datasets with different coincidence degrees within the effective range of ICP algorithm.(2)A new method called Four Initial Point Pairs(FIPP)is proposed which is based on initial four point pairs.The fast point feature histogram(FPFH)is used as the descriptor to extract salient feature points and a construct candidate point set by calculating the FPFH differences of points.On this basis,the feature points with the same name between two point sets are extracted.The similarity,equal distance,and position-constrained constraints are used to search for four pairs of points,and then the new pair of points satisfying the conditions are added continuously until the number of pairs meets the requirements of registration.Finally,the Total Least Square(TLS)algorithm is used to solve the rigid transformation matrix to complete the point cloud distribution.According to the experimental results of five different point cloud datasets,the method of FIPP has high efficiency and accuracy.(3)On the basis of FIPP,A faster point cloud global registration method is proposed based on point cloud RGB information.First,the number of RGB values in the point cloud dataset is counted,and colors having fewer points are removed by color filtering.Then,a candidate set of source datasets is constructed by comparing the similarity of the RGB values in the two datasets,and an improved FIPP algorithm is used to search for the initial four point pairs to complete global registration.Finally,local registration is achieved by means of the ICP algorithm.The results of two kinds of datasets experiments show that this method can achieve faster speed than FIPP algorithm while maintaining high accuracy.However,this method is suitable for point cloud registration with RGB values.(4)Drawing on the idea of region growing,an automatic method for extraction of urban road is proposed based on Mobil Laser Scanning(MLS)data.In this method,the initial seed is selected under the constraints of Gaussian curvature,elevation and the number of neighboring points,and then the decision of road growth is set based on the angle between the tangent plane of the seed point and the neighboring point.In addition,strategy under the multiple discontinuous roads in a dataset is also given.The method is well verified in five different types of roads,and its Kappa coefficient is above 90%.(5)Based on the idea of “step-by-step separation,layer-to-layer judgment”,a method of automatic extraction of urban street trees is proposed.The method is based on statistical filters to implement outlier elimination,horizontal projection and grid segmentation to achieve ground point segmentation,and the combination of point cloud clustering algorithm and normal difference to complete the separation of building points and build a cluster of candidate roadway tree points.Then,the candidate points clusters are stratified by constructing a voxel grid,and the residual elements are eliminated by analyzing the bottom-up cross-section change of ground points,low ground objects,and road poles in the candidate point clusters.Finally,the identification of a single street tree is done by finding the critical line of the cross-crown.