Font Size: a A A

Building Segmentation Method Based On Point Cloud Data In Multiple Scenes

Posted on:2024-09-05Degree:DoctorType:Dissertation
Country:ChinaCandidate:Z H SuFull Text:PDF
GTID:1522307301977129Subject:Information and Communication Engineering
Abstract/Summary:PDF Full Text Request
As the most important component of a city,building plays a crucial role in the threedimensional modeling of the city and the establishment of a digital city.Point cloud data is an important data source for 3D modeling,which can truly reflect the structural position and morphological information of objects.Li DAR scanners,depth cameras,and photogrammetry are means of obtaining three-dimensional spatial information,which can accurately and efficiently obtain three-dimensional point cloud information of buildings.We start from the perspective of the difficulty of building segmentation in different scenes.The research content covers the identification of various faces of a single building,extends to large-scale semantic segmentation of buildings,and further delves into the identification of roofs and facades of buildings in the shade of trees in complex scenes.The research content and conclusions are as follows:1.Individual building plane segmentation based on point clouds.A plane segmentation algorithm that combines the region growing algorithm with the distance algorithm based on boundary points is proposed to address the difficulty in identifying different objects in the same plane and common points between adjacent planes.Firstly,the building plane is segmented coarsely using the region growing algorithm.The uncoarse extraction plane points are obtained after removing the coarse extraction plane points from the original building points.Then,calculate the normal vector of the coarse extraction plane points and subsequently project these points onto the tangent plane.The coarse extraction plane’s boundary points are then extracted using the maximum value of the angle between the vectors formed between a point and its adjacent points on the tangent plane.Finally,the common point data of adjacent planes intersecting can be obtained from the distance threshold from the uncoarse extraction plane points to the coarse extraction plane’s boundary points.The optimal plane segmentation is obtained by combining the coarse extraction plane points with the corresponding common points.The optimal distance thresholds using the proposed method from the uncoarse extraction plane points to each plane boundary point of Cottage and Pantry were 0.025 m and 0.030 m,respectively.The highest correct rate,the highest error rate,and the F1 score value of the Cottage’s(Pantry’s)plane segmentation using the proposed method under the optimal distance threshold were 99.93%,2.30%,and 97.56%(98.55%,2.44%,and 95.75%),respectively.The proposed algorithm solves the problem of common points where adjacent planes intersect that cannot be extracted by the region growing algorithm,as well as the problem of plane over-segmentation caused by the RANdom Sample Consistency algorithm.2.Semantic segmentation of large-scale buildings based on point clouds.A deeper network structure,High Precision Range Search(HPRS)network,which accurately captures point features,is proposed to address the problem of the Point Net model’s weak ability to extract local features from point clouds.We divide the data into overlapping small areas,and then use the HPRS network to process the data from each small area.HPRS network consists of the set abstraction layer,the feature propagation layer,and the fully connected layer,with each set abstraction layer consisting of three parts: a sampling layer,a grouping layer,and an improved Point Net layer.Firstly,the set abstraction layer downsamples the data from the small region and performs high-dimensional feature extraction of the local region,consisting of the downsampling point as the center point and its adjacent points.Then,the feature propagation layer restores the original data by upsampling the data from the set abstraction layer and aggregates the features of each point.Finally,the features of each point are processed through the fully connected layer to achieve the semantic segmentation of the entire point set.Under 6-fold cross-validation,the OA,m Acc,and m Io U of the HPRS network are 84.70%,72.71%,and 61.35%,respectively.The network proposed can obtain local feature information that Point Net network did not collect,and can obtain more detailed local feature information than Point Net++ network.The network achieves superior performance on the S3 DIS dataset,with a m Io U declined by 0.26% compared to the state-of-the-art DPFA network.In the case of insufficient feature points within the ball,the HRPS network can obtain superior segmentation effect by using the adaptive ball query algorithm to form the local region composed of downsampling point as the center point and their adjacent points in comparison to the common ball query algorithm.In comparison to using either the mean pooling or the max pooling alone,the HRPS network can improve the segmentation effect by combining the two pooling functions in the improved Point Net layer to aggregate point features of the local regions.3.Building point cloud extraction in the shade of trees.A high-precision building point cloud extraction method based solely on 3D coordinate information of points is proposed to address the difficulty of accurately extracting building features due to the close adhesion between buildings and vegetation in complex scenes.Our method is divided into two stages: coarse extraction and fine extraction.In the coarse extraction stage of building points,our proposed method separates non-ground points from the original points using the cloth simulation filtering(CSF)algorithm and uses the region growing algorithm to obtain coarse extraction of building points from non-ground points.In the fine extraction stage of building points,considering that the region growing algorithm may fail to include the boundary points of the buildings,and the CSF algorithm may misinterpret facade points near the ground as ground points,the complete building points are obtained by combining two overlapping subsets(the building points and the facade points near the ground)extracted based on mask polygons.Our proposed method obtains the corresponding mask polygons based on the coarsely extracted building points by combining the Alpha Shape algorithm and neighborhood expansion method.Due to the possibility of adding certain tree points to the obtained building points,we use the region growing algorithm and the Euclidean clustering algorithm to filter out some discrete tree points from the building points to obtain optimized building points.Finally,the radius filtering algorithm is used to filter the discrete points obtained by removing duplicate points from the combination of the optimized building points and the facade points near the ground to obtain complete building points.The separation of facade points and roof points is achieved based on the normal vector threshold in the Z direction.The proposed algorithm has high extraction accuracy for Urban-Li DAR and Vaihingen.The proposed method performed outstandingly in the roof extraction of Vaihingen,especially since its precision was 20.73% higher than that of the Point Net network.The F1 score of the roof of Vaihingen using the proposed method was only lower than 0.28% for the HDLJME-GGO network.The proposed method maintains high accuracy in extracting building points,even in scenes where buildings are closely adjacent to trees.
Keywords/Search Tags:building, point cloud data, plane segmentation, semantic segmentation, extraction
PDF Full Text Request
Related items