Reconstructing 3D Buildings For Geographic Information Systems:An Image And DNN Based Approach

In the era of Big Data,3D digital maps and GIS are widely used in geographic information analysis and urban computing.However,the production cost of digital maps is expensive,especially that of 3D modeling of buildings that are important geographic objects,leading to the acquisition of high-precision 3D digital buildings map has a huge financial difficulty.As the map area increases,the efficiency of 3D building reconstruction is difficult to meet the demand.In this thesis,in order to solve the problems of high cost and low efficiency of 3D building reconstruction for GIS,we propose a fast and low-cost method for 3D modeling of buildings by using building images provided by Internet digital maps from the following four aspects.(1)We render the digital building maps in the browser through the API of the digital maps" developer platform and obtain images of 2D and 3D digital buildings maps through screenshots.These images are split and merged ensuring the border integrity of buildings.(2)To detect the building bottom contours in the 2D building image,firstly,we design the split and merge algorithm of complex buildings for handling complex buildings.Then the building bottom contours are detected from 2D buildings image via edge detection and contour recognition.Corner detection and edge smoothing algorithms are designed to eliminate the noisy points of the bottom contours of buildings,leaving only the vertices.(3)To calculate the height of buildings,we first take DNN-based object detection and instance segmentation to split a single building from 3D building images and calculate the pixel height of the building facades.Then a mapping model of building pixel height to metric height is established via a simple neural network to calculate the metric height of buildings.After that,an algorithm for calculating the height of occluded buildings is proposed.(4)Based on the above research content,we design and implement the software of 3D building reconstruction for GIS.We model 22,840 buildings in the urban area of Nanjing and take Amap and a 2m-resolution digital map as the baseline to evaluate the modeling accuracy,which contains three aspects:first,we evaluate the error of the building bottom contour based on the ratio of the contour area difference and the offset distance of sampling points.Second,the accuracy of modeling the heights of the buildings is evaluated.Third,we analyze the causes of errors.The results show that image-based 3D building modeling is able to provide acceptable-precision 3D geometry information of buildings.The above research shows that the scheme of 3D building reconstruction for GIS proposed in this thesis provides an effective method for fast and low-cost 3D reconstruction of large-scale buildings.