CityGML Building Model Texture Mapping Based On Multiview Coplanar Extraction

At present,the technology of 3D geometric model reconstruction in large streetscape maps is developing rapidly.With the advantages of low cost and short cycle time,texture mapping technology is gradually becoming one of the important technical means for 3D city modeling.At present,most online maps City GML of 3D building models lack texture information.The texture mapping of City GML 3D building model by crowd-sourced images can reflect the update of changing features of urban buildings visually,comprehensively and dynamically.To a certain extent,it can make up for the shortage of single data source photogrammetry 3D modeling,and the application advantages are outstanding.Unlike the methods of texture mapping using UAV images and Li DAR data,the crowdsource images vary widely in height,orientation,attitude and focal length.At the same time,the image sources are widely available,and the acquired images have problems such as huge quantity and multiple elevations.Due to the advantage of rich texture information in crowd-source images,high automation and fast extraction are important for the update of urban building fa?ade textures.Therefore,texture mapping of crowdsource images puts forward higher requirements for reliable,efficient and automatic data processing.Single-source texture mapping can no longer meet the needs of 3D reconstruction.This paper aims to improve the massing,automation and efficiency of texture mapping of City GML 3D building models.This paper conducts research on the retrieval of associated images,occlusion detection and fa?ade extraction of target buildings.The research work mainly includes the following points.(1)A method for retrieving and filtering building-related crowdsource images based on coordinate attribute information and angle constraints is studied to address the problem of low automation of image retrieval associated with the target building model.The method retrieves images associated with each elevation of the target building using the coordinate information recorded by the crowd source images.The online map is used to retrieve the images associated with the target building to find the images associated with each facade.Finally,the dot product formula constructed in this paper is used to filter the candidate images with distorted and blurred fa?ade textures.This method can provide texture sources for texture mapping of each elevation of the building model more automatically,reliably and efficiently using the crowd source images.(2)To address the problem of occlusion on building model facades in complex backgrounds,we investigate the use of a lightweight deep neural network Nano Det target detection model to detect vehicles,pedestrians,and other occlusions.At the same time,we combine gamma-enhanced visible green leaf index filtering vegetation to achieve intelligent extraction of candidate textures on building model fa?ade.Since the multisource multi-view images contain multiple texture information of each building fa?ade,some of the occluded areas can be repaired and image quality improved by using multi-view local single-shoulder transformation to repair the occluded or blurred areas,thus improving the completeness and clarity of texture information of texture mapped 3D building model fa?ade.(3)To address the problem of uncorrelated texture information mixed with candidate textures due to geometric shape differences in building facades of multi-view images from multiple sources,we investigate the construction of a high-quality mapping method for building model facade textures by combining texture source point-line multi-primitive features,geometric transformation single-response matrix and topology.The method extracts point-line co-planar features through multiple single-response matrix constraints to characterize the target fa?ade texture of multi-view surface images in two-dimensional space.At the same time,the method combines geometric topology to define the initial quadrilateral of the fa?ade,and uses Hough transform and iterative least squares to further optimize the fa?ade corner point positions,which can effectively improve the robustness of building fa?ade detection.Experiments show that the method in this paper can achieve building texture mapping based on crowdsource images without the support of extra-image orientation elements,i.e.,it is not subject to the geometric constraint that photogrammetric texture mapping requires co-linear equations.In addition,the method can effectively improve the quality of building model elevation textures,remove local texture blurring and geometric distortion,and is applicable to City GML 3D building model texture automatic fast update optimization.