Research On 3D Building Semantic Modeling From Point Clouds Based On PointNet Model

3D building models are important parts of geospatial information and the key to forming a "digital city",these models have been widely used in different areas such as urban planning,3D navigation,and cultural heritage protection.With the continuous development of lidar technology,point cloud data,especially airborne lidar scan data,has become the third important spatial-temporal data after vector maps and image data.The information contained in point clouds has unparalleled superiority of two-dimensional maps and images.Currently,3D building reconstruction from point cloud data is an active research topic in the photogrammetry,computer graphics,computer vision and remote sensing communities.However,current data-driven and model-driven methods focus on the geometric shape of building roofs,and the semantic relationship between the extracted planes is not clear.This thesis proposes a parametric modeling method that combines building roofs and facades,and uses deep learning technology to provide reliable prior information for 3D semantic modeling.The main contents are as follows:(1)Using deep learning technology to identify building roof categories.Traditional 3D modeling methods can only obtain partial geometric information,lacking semantic information and complete building category information;this article combines the data characteristics of building point clouds to improve the deep learning model and use the model to extract building roofs Category information;experiments show that the model in this paper has good generalization performance and recognition accuracy.(2)Designing a set of single and composed roof primitives,and representing these primitives by different parameters.Common primitives and primitive library designs only consider the geometric relationship between planes,ignoring global characteristics of the building.This thesis uses a completely top-down method that is based on the roof category of the building to design primitives,not only can It expresses the geometric characteristics of the plane and between the planes,and can also express the semantic information of each plane completely.After the semantic information is extracted and aggregated,CityGML is used for geometric and semantic integrated expression.(3)Improving the algorithm of parameter estimation.This thesis analyzes the descent process of the basic loss function,proposes an improvement strategy for the weighted piecewise loss function,and introduces the theoretical basis for automatically determining each parameter.Experimental results indicates shows that the weighted piecewise in this paper has a fast declining speed and good effect.