Research On Parametric Expression And Reconstruction Method Of 3D Building Facade Semantic Model

With the technological innovation in the fields of surveying and mapping science,information science and computer science,digital city and smart city develop rapidly with the help of 3D building models.In digital city,3D building models are usually organized in surface models and only with real texture.However,in smart city,3D building models with rich geometric and semantic information are needed.In this context,as being a key component of building,reconstructing semantic facade is of great importance.Unlike surface models for visualization,semantic facade models need to include more information,such as the location and label of each facade element,and even the geometric and topological relationships between different facade elements.Hence,compared with basic semantic segmentation task for relatively simple natural images,it is more difficult to reconstruct semantic facade models and extract elements’ relationships from images or point cloud data.At present,there are two challenges in the field of automatically reconstructing semantic facade models.The first one is the weak practicability of data-driven methods of semantic entity extraction(even the most advanced deep learning methods),which is caused by the unreliable data quality,such as occlusion,local missing and noises.The other is the complexity and diversity of facade structures,and these limit the generalization of facade structure description methods.Aiming to close above gaps,in this paper,the characteristics of facade structures are deeply explored,and a new method-facade layout graph-is proposed to better describe facade structure.Different with general grammar-based methods,the facade layout graph is a hierarchical graph model based on compound graph,which can describe the geometric and topological relationships between different facade elements.Hierarchical graph can fully and directly represent the layout of facade structures with strong interpretability and straight expression form,making it more readable and can better support the reconstruction of semantic facade models.Then,the parameter optimization of facade layout model using Bayesian framework is used to automatically reconstruct semantic facade models in the way of gradual approximation.Experiment results demonstrate the reliability and efficiency of these methods.The key contributions of this paper are as follows:(1)Propose a new method,facade layout graph,for the description of facade structure.In this method,the layout of facade structure is considered based on compound graph model,the Principle of architecture form,and Gestalt theory.The floor layout is regarded as the top layout and the facade elements are regarded as bottom layout when building the compound graph model.The nodes and edges in this graph model have corresponding attributes to ensure the unique mapping relationship between the graph model and facade structure.(2)Study the parameterization method of semantic facade model.Based on the structure description parameters and construction parameters of the facade layout graph,the parameter set is used to represent each semantic facade model.The global constraints between different parameters are established by analysing the topological association between facade elements.The geometric constraints of parameters are established by analysing laws and regulations related to daylighting and energy saving.With the reference to these constraints,the value of every parameter can be limited into a proper range to ensure the effectiveness of this method.(3)Propose a new standard for the reconstruction of semantic facade model with level of details(Lo Ds)based on the parametric design of the facade layout model.In this new standard,semantic facade models are divided into three levels-F0,F1 and F2.The semantic facade model in F0 only contains basic facade layout.Facade elements in the F0 model have the same shape and size,and have the same layout on each floor.The semantic facade model in F1 and F2 contain more detailed information,especially geometric information of facade structures,such as the misalignment between floors,the inconsistent distribution of elements between different floors,and the different shapes of elements.The F2 model also contains the 3D information of facade elements.(4)Propose a new optimization method for automatic facade reconstruction based on the facade layout graph model and Bayesian framework.This framework is divided into two parts and can be applied to image data and point cloud data respectively.For the facade image data,instance segmentation is used at first to extract the geometric and semantic information of facade elements,and then EM algorithm is used to reconstruct facades.For point cloud data,since there is no existing reliable instance segmentation algorithm for them,the facade geometric information from point clouds is extracted by refining the facade structure.Then the random strategy is used to Optimization process.In experiments,3 kinds of point cloud datasets,1 kind of crowdsource image dataset and 3 large-scale urban areas are tested.The experimental results demonstrate that,compared with traditional grammar-based methods,the proposed methods in this paper have better accuracy and efficiency,and lower storage requirement.Furthermore,these methods can maintain the formal intuitiveness of facade structure description.