Classification Of Terrestrial Laser Point Cloud In Urban Scenes Considering Density Variation

3D(three-dimensional)terrestrial laser scanning technology is capable of acquiring accurate 3D spatial information and real shape of target surface in a non-contact manner,which has the advantages of flexibility,speed,high accuracy,strong initiative,and realtime acquisition.It is widely used in urban planning,3D modeling,unmanned driving and other fields,and has become a key data source for digital twin city construction,smart city construction,and urban informationization construction.Due to the scanning distance,angular resolution and other factors,terrestrial laser scanning data suffers from an uneven density distribution.In low-density areas far from the scanning center,the continuity of the point cloud space is poor,and it is difficult to restore local structures,which brings challenges to subsequent point cloud processing.To address the above problems,this thesis proposes a density-adaptive feature extraction method,which aims to improve the problem of point density variation in terrestrial laser scanning data.In this thesis,we focus on point cloud pre-processing,angular resolution estimation,point cloud feature extraction and classification in the process of point cloud processing,and the main research contents are as follows:(1)Point cloud pre-processing.Three common forms of point cloud index organization are compared and studied: regular grid,octree and KD-tree index.The appropriate index method is selected based on the different purposes of point cloud processing.In terms of point cloud filtering,three filtering methods: progressive TIN(Triangulated Irregular Network)densification,point cloud filtering based on point density analysis,and cloth simulation filtering are compared.Finally,in order to reduce the influence of ground points on the subsequent processing and reduce the amount of points,cloth simulation filtering is used to remove the ground points in the experimental scenes.(2)The neighborhood analysis of randomly picked points.As an important indicator for measuring the accuracy of point cloud data,angular resolution is a crucial factor affecting the density distribution of terrestrial laser scanning data.It is usually regarded as a known parameter and directly brought into density-adaptive algorithms.However,in practical applications,due to limitations of the scanning instrument,transmission format,and other factors,the situation of unknown angular resolution is difficult to avoid.This situation is less considered in existing density adaptation studies,limiting the generality of the related algorithms.To address this problem,this thesis proposes an angular resolution estimation method,called the neighborhood analysis of randomly picked points.The algorithm calculates the angle between the central beam and the corresponding neighboring beams to construct a histogram.By counting the horizontal and vertical angles in different interval,the final estimate result of angular resolution is obtained.The comparison experiments on several datasets demonstrate that,compared with the point spacing-based method,our method has good angular resolution estimation accuracy,and good stability for different types of objects and parameter settings,moreover,it can provide data support for subsequent density-adaptive processing.(3)Classification of terrestrial laser point cloud taking density variation into account.To address the problem of uneven point density distribution of terrestrial laser point cloud,this thesis proposes a density-adaptive feature extraction method by analyzing the factors affecting the density variation of terrestrial laser point clouds.The method improves the traditional projection density feature based on angular resolution at the feature design level,and combines commonly used geometric features to construct a set of densityadaptive feature combination for describing point cloud shapes,density distributions,geometric surfaces,etc.Finally,Random Forest classifier is used to classify and evaluate the feature importance,and the effectiveness of the improved projection density features is verified by comparison experiments.Through experimental validation on four datasets,compared with the traditional projection density,the method in this thesis can weaken the effect of density variation on feature extraction and improve the multi-objects classification accuracy,especially for small-size objects,such as pole,car and scanningartefacts.The overall accuracy of classification is improved by 3.2%,6.48%,3.11% and0.33% respectively.Our method can provide a new solution for the improvement of multiobjects classification accuracy of terrestrial laser scanning data.