| As a popular application technology,UAV oblique photogrammetry technology is playing a key role in urban and rural large-scale landform survey,geographic information collection,disaster detection and other fields.This technology has a high degree of automation in the work of UAV oblique data acquisition and computer graphics processing.However,the work related to the layout design of image control points and the prick point on the internal map still needs artificial subjective vision,and the degree of automation is low.Therefore,to make better use of oblique UAV photogrammetry in civil engineering and increase operational efficiency,this paper proposes a method that can automatically plan the layout scheme of image control points according to different parameters,and an automatic identification method of ground control points based on deep learning technology.By improving the efficiency and accuracy of field layout and interior ground control points,the accuracy of three-dimensional model is improved,the operation time is shortened and the production cost is reduced.The main research is divided into the following points:1)The structure and measurement principle of UAV tilt photogrammetry system are introduced.This paper describes the data acquisition and processing process of UAV tilt photogrammetry,including the internal and external work flow and related specifications of conventional tilt photogrammetry.Taking the actual UAV image data as an example,this paper introduces the core technology of UAV tilt photogrammetry,and obtains the image matching results of Nanmen Building of Anhui University of Science and Technology and the monomer model of the national key laboratory.2)The norms and specific methods of field layout planning of image control points are studied.The advantages and disadvantages of the three methods are compared,and the problems of traditional manual layout of image control points and indoor pricking methods are summarized.According to the relevant technical principles,the high-altitude flight route of the UAV is calculated.On this basis,a method for automatically obtaining the position image of the image control point is proposed.Three image control point distribution schemes with a small number of centers around and around and automatic design were used to process and produce orthophoto images and 3D models in a rural experimental area.The experiment shows that,on the basis of the same set of UAV oblique photography source data,in Area A,the mean square error of the result plane processed by the automatic layout scheme is 0.08m,and the mean square error of the elevation is 0.10m.Compared with the method of adding a small amount of intermediate layout around and around,the mean square error of the elevation is increased by 0.09m and 0.04m respectively,and the mean square error of the elevation is increased by 0.06m and 0.05m compared with the method of adding a small amount of intermediate layout around and around.The mean square error of the results processed by the automatic layout scheme in region B is 0.08m,and the mean square error of the elevation is 0.11m.Compared with the method of adding a small amount of intermediate layout around and around,the mean square error of the elevation increases by 0.08m and 0.06m,respectively.Compared with the method of adding a small amount of intermediate layout around and around,the mean square error of the elevation increases by 0.06m and 0.01m.The accuracy of orthophoto map meets the requirements of large scale topographic mapping specification.3)The image corner recognition algorithm and the target detection algorithm based on deep learning are analyzed.The UAV image enhancement algorithm based on color space is proposed.The original UAV image is denoised and enhanced.The image enhancement is compared with the single-scale Retinex algorithm and the multi-scale Retinex algorithm in the dark background.The results show that the improved algorithm is better.Based on this,the SSD+LSD enhancement processing algorithm is proposed.The selected image is trained by deep learning 300 enhancement algorithm.Combined with line segment detection algorithm,the shape angle feature threshold of image control point image is set to identify the geometric center mark of the selected image.The experimental results show that compared with the traditional SSD300 algorithm and YOLO algorithm,the accuracy of image control point recognition is improved,and the accuracy can reach 94.75%.The positioning accuracy of image control point can reach three pixels.4)Taking the new campus of Anhui University of Science and Technology as the experimental area,the image control points on the campus are automatically arranged,and the automatic identification of image control points is used to assist the pricking work in the internal processing.The experiment shows that compared with the traditional manual layout of image control points and puncture of image control points,the layout and puncture points with high degree of automation can improve the production efficiency of the three-dimensional model,and the obtained three-dimensional real campus model meets the level I standard.Figure[53]Table[12]Reference[101]... |
PDF Full Text Request