Robust Building Modeling Method Based On Image Dense Matching Point Cloud

Building 3D model is the most important basic data of urban model,which plays an important role in urban operations,rescue guidance,urban planning and construction,intelligent transportation and other fields.The traditional artificial modeling method is inefficient and costly.The surface reconstruction method based on grid cannot operate single buildings and the model surface constructed is not smooth enough.The modeling method based on point cloud has become the mainstream.However,the existing method based on point cloud has the problems of single building reconstruction model shape and group building can not be refined modeling.To solve these problems,this thesis focuses on three key technologies,i.e.,hybrid modeling of prior hypothesis and dial-and-conquer strategy,point cloud completion,and point cloud segmentation and model fusion,in order to improve the robustness of building shape in single building modeling and the precision of model in group building modeling.The main research contents of thesis are as follows:(1)Modeling method of single building based on a mixture of prior assumption and divide-and-conquer strategy.The main method of urban building modeling is based on Manhattan hypothesis,but this kind of method can only build multi-level flat roof building model,can not build other shape building model.In this thesis,based on the idea of prior assumption and roof point cloud segmentation technology,an improved modeling method for single buildings based on Manhattan hypothesis is proposed.In addition to building models with flat roofs,Manhattan models of buildings with other shapes can be constructed,which improves the robustness of the modeling method to building shapes.In order to further improve the modeling accuracy of sloping roof buildings and balance the modeling efficiency,based on the idea of divide-and-conquer strategy modeling,this thesis introduces the method of intersection plane selection on the basis of Manhattan hypothesis,modeling the main body and roof of the building separately,achieving the balance of modeling accuracy and efficiency.(2)Group building refined modeling method for incomplete point cloud.At present,the group building modeling method is to integrate the whole building group,which can not achieve the effect of fine modeling.Based on point cloud block and model fusion,this thesis conducts monomer processing for each building in the building group.Aiming at the absence of neutral point cloud in the building group,adjacent planes are used to complete the building group and realize the fine modeling of the building group.Using this method,the Manhatton model and the original inclined roof model of group buildings can be well constructed on the field spot cloud generated by Blend MVS dataset.(3)Actual scene experiment verification.In order to verify the effectiveness and practicability of the proposed building modeling method,this thesis conducted data collection based on the UAV platform independently built by the research group,and completed the building modeling experiment in the actual scene.The average error of the closest distance of the actual scene model is less than 0.1m,which achieves the accuracy equivalent to the current best method,and improves the robustness and efficiency of modeling.Experimental results verify the feasibility and effectiveness of the proposed method.