Research On Structure-preserving 3D Building Simplification Methods

Drone-based photography technology can automatically generate a large-scale 3D urban surface model,which has become a common method for quickly acquiring 3D spatial data,and is widely used in many fields such as digital cities,transportation,city planning and etc.However,the 3D mesh model generated by the photogrammetry technology has some problems,such as huge number of vertices and high noises,which brings difficulties to data storage,transmission and semantic expression.Building simplification,that is,to represent the original 3D building with a regular model composed of a small number of triangular faces,is the main way to solve the above problems.A building usually has strong appearance structural features,including roof contours,walls and corners,door and window eaves,etc.In order to retain the appearance structure of the original 3D model,this paper extracts the simplified model of the building in layers to make it closer to reality.The main research contents include:(1)We proposed a regularized reconstruction method for 3D building models based on roof contours,which can convert the complex and noisy 3D building grid models into simple,regular building models.Firstly,we segment the roofs and extract their outer-contours,and further refine the rough outer-contours;then,we extract the roof inner-contours by an improved plane fitting method;thirdly,we combine the inner-contours and outer-contours,and generate planar primitives by an optimization with coincidence constraints and collinear constraints;fourthly,the non-planar part of the roof surface is simplified through a connectivity clustering;finally,we restore the elevation information of the vertices in the plane model by considering the elevation constraints and coplanar constraints,and further restore the walls and bottom surfaces to get a simplified model for the main structure of the building.(2)We proposed a door and window reconstruction method based on deep learning instance segmentation,which can simplify the structure of the facade door and window details and make the building model closer to reality.Firstly,the trained Mask R-CNN framework is used to predict the doors and windows in the building facades.For the problem of inaccurate boundaries of the predicted doors and windows,an improved maximum inter-class variance method is proposed to search the four sides to determine the accurate boundaries;we evaluate the scale similarity and texture similarity among doors or windows and group them for regularization,so that the doors or windows in the same group have the same scale,and keep horizontal or vertical parallel;finally,we project the optimized 2D doors and windows to the 3D simplification building model,which is integrated with the building model through a 3D Boolean operation.The simplified method for 3D buildings proposed in this paper has the characteristics of preserving the building structures.Through a large number of experiments,it is shown that our methods in this paper can better achieve the structure-preserving simplification of most buildings,retain the structural characteristics of the original 3D models at different levels,both of which are close to the original 3D model.Therefore,the proposed methods in this paper can effectively regularize and reconstruct the 3D building model,and has a promoting effect on compressing model data,reducing model noises,which increases its industrial application ranges.