Research On Point Cloud Labeling And Building Extraction Based On Campus Scene Layout Semantics

The construction of digital campus and intelligent campus is not only an important content of "Educational Informatization 2.0",but also an important foundation to improve the level of campus management and service.The traditional modeling and data acquisition methods based on image and manual measurement have slow speed and high cost,which limits the development of intelligent campus construction.At present,the aerial images of the campus are quickly obtained by UAV and densely matched into point clouds,thus modeling the campus has become an important method to speed up the construction of intelligent campus.However,there are some problems in the large-scale point cloud acquired by UAV,such as the difficulty of extracting campus architecture in complex environment,the lack of campus large-scale point cloud data set with category label obtained by dense matching technology,and the difficulty of labeling campus point cloud with points of more than 10 million,which brings difficulties to further campus building extraction and intelligent campus automatic modeling.How to label the point cloud quickly and accurately,automatically extract the building point cloud in the complex environment,and construct the basic space framework of the intelligent campus is an important means to apply the interdisciplinary method to speed up the construction of the intelligent campus.it is also an important research content and direction of educational information construction.In order to solve these problems,this thesis focuses on the theme of unmanned point cloud data labeling and building automatic extraction in the process of intelligent campus construction.The main research contents include:(1)This thesis summarizes the labeling methods of point clouds,the open point cloud data sets and the methods of semantic segmentation of point clouds,and focuses on introducing and comparing the algorithms that can be used to deal with outdoor large-scale point clouds,and according to the characteristics of campus point clouds,select the appropriate algorithm for the later semantic segmentation experiment.(2)In this thesis,a data acquisition scheme of using UAV to quickly obtain campus point cloud is constructed.on this basis,the idea of campus scene layout semantics is introduced to label campus point cloud data,and initially distinguish architectural point cloud from non-architectural point cloud.Then,according to the format of Semantic3D data set,a standard campus point cloud data set is constructed.(3)The point cloud semantic segmentation experiment is carried out on the constructed campus point cloud data set by using SPG and RandLA-Net algorithms,and the accuracy is tested by the verification set while training on the training set.finally,the optimal model is used to automatically identify and extract the buildings on the test set,and we obtain a good experimental result,which shows that the semantic segmentation model can accurately extract buildings from the campus point cloud.The innovation of this thesis is mainly reflected in the following aspects:(1)A method of point cloud labeling based on multi-source data such as campus layout semantics is proposed,which significantly speeds up the labeling of point cloud in campus buildings.(2)We construct a standardized campus dense matching point cloud data set,which makes up for the lack of this kind of data set.(3)The SPG and RandLA-Net algorithms are used to train on the self-made data set,and the semantic segmentation model which can accurately extract the building point cloud is obtained,which lays a foundation for the unitization standard process of campus building point cloud.