A Semantic Modeling Framework-based Method For 3D Building Reconstruction From Airborne LiDAR Point Clouds

In recent years,the semantic information is more and more important in the city modeling field.However,although the existing methods such as modeling with triangulated irregular network,have a high automation and precision,there is little semantics information preserved.Nevertheless,some methods such as grammar based methods that can provide abundant semantic information require many manual intervention,so they are inefficient.As the main component of a city,building has complicated structures and various architectural styles,that makes the semantic reconstruction of the building a challenging task,because that refers to many key technical such as building segmentation,architectural rules detection and semantic feature recognition.In this paper,based on the summery of previous works,we analyze the problems and limitations in the current building reconstruction field.To improve the existing methods,we choose the airborne LiDAR data as our data source,and base on the airborne LiDAR data's feature and model driven modeling concept,we present a semantic modeling framework-based method for building reconstruction.In our method,a semantic modeling framework is designed for the building's semantic description and modeling.Additionally,the extraction of the building semantic information is researched,some related algorithms for recognizing the semantic features and extracting the information are presented.Finally,the technical feasibility of our method is validated using a group of the airborne laser scanning data sets,which indicate that our method can simplify the reconstruction process from a point cloud and generate 3D building models with high accuracy and rich semantic information.The main research works are illustrated as follows:(1)In this paper,we explore the new demands from the various industries,such as building segmentation,rich semantic information,precise geometric details and high automation.And after analyzing the features of different LiDAR platform,including the airborne LiDAR,mobile LiDAR and ground based LiDAR platform,we decided to select airborne LiDAR data for our research.(2)Analyzing the advantages and disadvantages of the existing methods,such as the data driven methods,model driven method or grammar based methods and knowledge based methods,a comparison is made through several aspects involving the data source dependency,model quality,semantic information richness,automaticity and sharp feature preservation(3)Aiming at the shortcomings of the existing methods,a semantic modeling framework for 3D building reconstruction is proposed.After analyzing and comparing the existing City modeling methods,we propose a lightweight semantic modeling framework for 3D building reconstruction,and elaborate on its design concept and architecture.The framework consists of two parts:the extensible Building Modeling Language(XBML)and the corresponding building component library.XBML is responsible for the semantic and parametric description of the building structure,and the corresponding building component library encapsulates the concrete procedural modeling algorithm,and generates the concrete3D model according to the parameters provided by the XBML tag.The framework separates the semantic description from the geometric modeling of the building,allowing the modeler to focus more on the description of the building without caring about the details of the geometric modeling,it also make the modeling process object-oriented,Readable,reusable,thus the framework simplifies the modeling process.(4)For the roof outline extraction,two algorithms are proposed:the a-shape based outline extraction algorithm,outline extraction algorithm based on combination of bilateral filtering and Gaussian mixture model.The first algorithm applies to most of the roof point cloud which is scanned well,and the second algorithm is mainly for the roof boundary which is scanned incompletely.(5)To reconstruct the pitched roof,we propose a shortest cycle-based method for the pitched roof reconstruction.Through the improvement of roof topological graph(RTG)algorithm,the pitched roof reconstruction can be seen as the classical problem about the least closed loops searching,which improves and simplifies the existingtheories.(6)A floor analysis algorithm based on point confidence is proposed.The algorithm can utilize the distribution characteristics of the boundary point of the doors and windows in the facade point cloud to detect the floor structure and extract the semantic information.Instead of using the covariance matrix to evaluate the point confidence,the algorithm proposed in this paper use an improve way to evaluate it,so our algorithm has a lower complexity in space and time than the previous works.(7)Aiming at the problem that the traditional reconstruction methods are not concerned about the detail structure of the building,this paper puts forward the algorithm for the roof detail recognition,and realizes the detection and size estimation of the eaves and the parapet.(8)Based on the boundary point confidence information of the doors and windows evaluated during the floor structure analysis,this paper presents a window detection algorithm based on point confidence.Through the boundary point detection of the windows,the location of the windows,the boundary points of the windows can be extracted roughly.And base on this,an eight-quadrant boundary point detection algorithm is adopted to realize the accurate extraction of doors and windows.(9)Research on the transformation from the semantic information to semantic modeling framework.To preserve the hierarchical relationship between semantic information and facilitate the transformation of the semantic information to XBML,this paper studies the hierarchical storage of the semantic information,and uses the object relational database to store the semantic information,using its object-oriented feature to describe the extracted semantic information and hierarchical information.(10)Research on the design and implementation of building modeling language-XBML parser.One of the difficulties in the framework is how to parse XBML into an build-in program language supported by the 3D modeling platform.This parsing process is not just the process of XML parsing,actually,it is a parsing process from XBML to another programming language(MAXScript,Python).In this paper,some key technical in the parsing process are elaborated,including the parsing of elements and general attribute,parsing of the parent elements and sub-elements,parsing of the attached attributes and type conversion involved in the process of the attribute parsing.