3D Building Reconstruction Based On Crowdsourcing Images

3D city modeling is a hot topic in long-term research in the fields of photogrammetry and computer vision.Compared with traditional two-dimensional geographic data,the three-dimensional city models carry more information and has been widely used in many fields such as municipal planning,city decision support,electronic communication,aviation and so on.Buildings are the most important constituent elements in a city and the key to a three-dimensional city model.In recent years,with the rapid popularization of mass shooting equipment such as smartphones,tablets,consumer-grade digital cameras,etc.,the method of data collection has become increasingly popular.The crowdsourcing images captured by these non-professional equipments have attracted increasing attention,providing a wider source of data for the three-dimensional reconstruction of buildings.How to carry out 3D reconstruction of buildings from crowdsourcing images taken by non-professional equipment is a very meaningful and challenging task.In this regard,this paper studies the use of non-professional crowd source images for precise three-dimensional reconstruction of building facades with precise geographic positioning.Specifically,the research content of this article mainly includes the following three aspects:(1)Complete reconstruction of the building facades from the crowdsourcing images based on structural characteristics.When generating a building point cloud from crowdsourcing images,the connection of the multi-view image matching is often disconnected due to various reasons,resulting in the failure to generate a single building facade point cloud that completely contains the various facades.This paper takes use of the spatial relationship constraints of different facades implied in the structural characteristics such as the Manhattan world hypothesis,the similarity of the facade structure,and the constraints of the building outline,and gradually restores and reconstructs the spatial relationship between the fracture point clouds to reconstruct a complete building facade point cloud that contains all facades.(2)Geographic registration of fa(?)ade point cloud based on open LiDAR data.Accurate geographic localization of the building facade point cloud generated by crowdsourcing images has always been a problem to be solved.This paper studies the method of accurate geographic registration by registering building image point cloud and open LiDAR data obtained from open geographic data website.In order to avoid the misregistration caused by the repeated structure of the building,we propose a normal constraint coherent point drift algorithm.The proposed method for geographic registration of image-based building point clouds can reach an accuracy of decimeter level.(3)Regularization of windows point cloud on facades using semantic and geometric information.In view of the insufficiency of accuracy and regularity of the windows point cloud in the fa(?)ade reconstructions,this paper studies the use of the deep learning model,which has strong feature learning capabilities,to extract the windows on the facade of the building from the images.And then windows are regularized on the generated orthophoto depth map to model the building facade concisely.