Research On Regularization Of Building Contour Based On UAV Image

As a new flight platform,UAV has gradually become a new favorite in the field of photogrammetry with the advantages of high flexibility,fast speed and easy operation.Nowadays,UAV photogrammetry gradually replaces the traditional measurement and is widely used in small area large scale mapping.In the UAV large-scale mapping,it is still necessary to manually extract the features on the image,which greatly increases the mapping cost.And the building is the most important part of the feature,Research on automatic extraction of buildings from UAV images can reduce labor costs and speed up work efficiency.Therefore,the automatic extraction of buildings from images has a very important research value.In view of the small amount of UAV image data,deep migration learning is used to extract buildings from UAV images.In this paper,the nonlinear Gauss Helmert model with orthogonal constraints is used to fit the extracted building contour.The main contents and research results of this paper are as follows:(1)This paper introduces the composition and principle of UAV aerial photogrammetry system,and introduces in detail the work before,during and after photogrammetry as well as the related important technologies.Combined with specific engineering cases,the layout and acquisition methods of image control points in UAV photogrammetry are discussed in detail,and the aerial images are processed based on contextcapture software cluster technology.Finally,three orthophoto images of the target area are produced through accuracy verification.(2)This paper uses semantic segmentation u-net network to extract buildings from orthophoto images.Aiming at the problem that there are too few data in the target area,the u-net network is trained in the open source data set by deep migration learning method,and then the buildings in UAV Orthophoto Image are extracted by migration learning.Experiments show that when the target data is missing,the method of deep transfer learning is used to extract the target data,and the extraction result is better.(3)For deep learning extraction of buildings,there are some problems such as jagged and irregular building contour.This paper analyzes and compares the existing regularization methods of building contour.In the corner detection method,the least squares is replaced by the global least squares to fit the building contour.For rectangular buildings,the nonlinear Gauss Helmert fitting model with orthogonal constraints is derived.Compared with the existing regularization algorithms,the experimental results show that:compared with the existing regularization algorithms,the regularization model proposed in this paper can make the rectangle orthogonal better and improve the boundary accuracy.Figure[31]table[14]reference[55]...