Rresearch And Application Of Deep Learning-based 3D Building Reconstruction

3D modeling of buildings plays a crucial role in geographic information systems and finds extensive applications in smart cities,panoramic maps,and wireless communication network optimization.However,conventional techniques such as oblique photography and laser point cloud scanning are time-consuming and expensive.Therefore,researchers are exploring low-cost and high-efficiency methods to obtain large-scale building three-dimensional geometry information.In this regard,this thesis proposes a practical approach that leverages building layer images get from online electronic maps to extract building bottom outlines through image processing technology.A deep learning method is then used to obtain the vertical edge height of the building,enabling fast,efficient,and cost-effective 3D modeling.The research work presented in this thesis is divided into four parts.(1)Modeling the building’s bottom outline involves extracting outline coordinates from the two-dimensional layer images of buildings provided by electronic maps.This is achieved through the use of image processing technologies such as edge detection,contour detection,and corner point detection.Complex buildings are split into multiple independent parts,then base polygons are modeled for the bottom outlines.(2)Acquiring the heights of building vertical edges is the main focus of this thesis.Three-dimensional building layer images provided by electronic maps are used to collect and annotate a large number of buildings and their vertical edges to construct a set of training data.A building vertical edge detection model is built using the YOLOV5 pretraining model and attention mechanism.A pixel height to metric height mapping model based on BP neural network is constructed through the batch generation of training data that represents the mapping relationship between building edge pixel height and metric height in multiple scenes.The edge detection model identifies the vertical edge of the building,and the mapping model changes its pixel height to metric height.An independent auxiliary grid mechanism is used to handle special cases such as edge occlusion and multi-height edge in the process of building edge recognition.(3)Building texture reconstruction based on crowdsourced images involves using a YOLO V5 model to construct a building facade feature recognition model.The model identifies building facade features,including the location and quantity relationship of doors,windows,and balconies,from the building images uploaded by Internet users.The 3D geometric model of the building is then refined and improved through steps such as repositioning and feature restoration.(4)The development and application of a 3D building reconstruction software is based on the research conducted in the aforementioned sections.This software is designed and developed for large-scale scenarios,and function tests and performance tests are carried out for practical application scenarios.In addition,the 3D modeling of large-scale buildings is carried out for typical areas of Nanjing’s dense urban area,in combination with a practical application project.The electronic map with 2m precision of Nanjing’s urban area is used as a benchmark to evaluate and verify the modeling accuracy.The evaluation includes three dimensions:bottom contour vertex sampling bias,bottom profile area error,and height error at different vertices,with multiple evaluation indicators used for each dimension.The evaluation results confirm the feasibility of the proposed building 3D modeling scheme.The above research work demonstrates that the approach of 3D building modeling based on image processing and deep learning proposed in this thesis can quickly,efficiently,and cost-effectively provide 3D building information with a certain level of accuracy using building layers of electronic map.This method has good application prospects in various fields such as smart cities and mobile communication network optimization.