Research On Building Extraction Based On TLS Point Cloud Data

In 2020,the State Council issued the Guidance Opinions on Comprehensively Promoting the Transformation of Old Urban Communities,and thus the city entered the era of rapid self-renewal,of which building information acquisition is the key.Traditional surveying and mapping technology is complicated,intensive,long,and incomplete,which cannot meet the requirements of building information acquisition at this stage.Terrestrial 3D Laser Scanning(TLS)technology has the advantages of fully automatic,non-contact,complete data acquisition,fast efficiency and high accuracy,which not only breaks the shackles of traditional measurement methods,but also has low cost and can be used in 3D modeling to achieve high reduction of data storage of actual objects.However,point cloud data has the characteristics of large quantity,covering many kinds of features,unstructured,etc.It is difficult to separate building point clouds from point cloud scenes,so the study of building extraction from TLS point cloud data can effectively promote the process of building information extraction,and provide scientific theoretical basis and practical proof for the subsequent point cloud processing research.In this paper,the research combined the research and algorithm results obtained by domestic and foreign scholars as well as the experimental scheme,and carried out the acquisition and pre-processing work of TLS experimental point cloud data,ground filtering experiments and the research on the classification and extraction of building point cloud data,and had haven achieved certain results.The main contents and results of the research in this paper are as follows:1.The characteristics and working principle of ground 3D laser scanning technology were explained,the information of Z+F IMAGER 5010 C 3D laser scanner was introduced,the experimental point cloud data acquisition and internal stitching,denoising,and extraction of thinning were performed for the data acquisition area of this paper,and finally,after cropping and manually adding recognition labels,the subsequent filtering experiment and the accuracy verification of building point cloud classification extraction were prepared.The data would be prepared for subsequent filtering experiments and accuracy verification of building point cloud classification extraction.2.Being comparing and analyzing the progressive encrypted triangular mesh filtering(PTD)algorithm with the fabric simulation filtering(CSF)algorithm.On the basis of elaborating the working principle,process and parameter settings of PTD and CSF algorithms,ground filtering experiments of the above two algorithms were conducted respectively,and the filtering results were analyzed by Kappa coefficient,which measures the classification accuracy,OA coefficient,and cross-tabulation for reasonable analysis.The experiments showed that the Kappa and OA values of PTD were 0.8301 and 0.8612,respectively,and the Kappa and OA values of CSF algorithm were 0.9228 and 0.9987,respectively,under the best filtering effect,and the class I error(miss-score error)of PTD is 0.1388,which was higher than the value of 0.0013 of CSF,and the class II error(The values of Class II error(misclassification error)were 0.0290 and 0.0774,respectively,and the latter is higher than the former and the overall value was lower.In order to filter out the ground point clouds as completely as possible to ensure the accuracy of subsequent building point cloud classification extraction,the filtering results of CSF algorithm were selected for the next step of building point cloud classification extraction research.3.Coarse extraction of building point clouds based on the fusion of DBSCAN clustering and point cloud section projection.The principle of DBSCAN clustering algorithm is introduced,and the cross-sectional characteristics of building,street light,vegetation and vehicle point clouds are analyzed in depth,and the corresponding threshold values are determined through experiments.Based on the relative independence of groundless experimental data,the point cloud clusters of buildings were coarsely extracted by fusing DBSCAN clustering and point cloud hierarchical cross-sectional projection.4.Based on the improved RANSAC combined with the area growth algorithm,the point cloud fine classification of vegetation and building confusion area is carried out.Since the spatial location of vegetation around the building is too close to the building,the confusion area is not effectively classified in the coarse extraction stage of point cloud clusters.Therefore,based on the review of the traditional RANSAC algorithm,the defects of it are analyzed,and the improved RANSAC algorithm is proposed for the problem of fusing the characteristics of laser point normal vectors on the wall,and then the confusion area is finely classified by combining with the area growth algorithm,and finally the The accuracy evaluation and building 3D model construction are finally performed.The results show that the average accuracy of building point cloud classification extraction is 93.13%,and the accuracy of building plane boundary length meets the relevant requirements and can reach the centimeter level.