Research On 3D Reconstruction Method Of Urban Real Scene Based On UAV Aerial Images

With the promotion of digital smart cities,the high-precision 3D reconstruction technology of urban scenes has become a research hotspot.Low-altitude UAV aerial photography,as an important technical means of acquiring images of ground scenes at present,has important application significance in such fields as urban construction planning and resource surveying.At present,domestic mainly use UAV images and adopt commercial software to carry out 3D reconstruction of urban scenes,and certain achievements have been made,but there are still problems such as time-consuming reconstruction,poor accuracy of 3D models and low granularity.To address the above problems,this thesis combines the characteristics of UAV images and deep learning algorithms to research 3D reconstruction technology,the main work includes:1)An incremental motion recovery structure algorithm is proposed.In order to improve the acquisition rate and accuracy of camera poses and sparse point cloud models,this thesis improves the traditional incremental motion recovery structure algorithm.First,in the feature matching stage,the GPS information in the EXIF attribute of the UAV aerial image is used to constrain the feature matching process,which effectively reduces the computational cost of feature matching.Then,a correlation measure combining the number of image feature points and the length of feature trajectory is proposed,through which optimal viewpoint selection and reliability region construction are achieved.Experimental results on both publicly available datasets and self-collected data show that the overall time consumption of the method in this thesis can be reduced by up to 49.11%and the reprojection error by 1.39% compared to the general open source Sf M software COLMAP.The method in this thesis can effectively reduce the time consumed by the incremental motion recovery structure algorithm on UAV images and improve the recovery accuracy of camera poses.2)A depth estimation network based on UAV aerial images is proposed.The network extracts multi-scale depth features of the image and performs depth estimation on different scale features in stages so that the depth range is constantly close to the true depth value,while using a cyclic coding and decoding structure to regularise the cost body and reduce memory consumption.Experimental results on the aerial photography dataset WHU-MVS show that the average absolute error of the method in this thesis is reduced by 1.73% compared to REDNet.The memory consumption of this method is only 9183 M for an image with an input resolution of 1152×864.The algorithm proposed in this thesis can reduce the computational memory consumption and improve the depth map estimation accuracy at the same time.3)A 3D real-site point cloud reconstruction and semantic segmentation method is proposed.In this thesis,a multi-rotor UAV equipped with five cameras is used as an image acquisition platform to collect an aerial sequence image dataset of a real campus scene,including 1700 high-definition images,by tilt photography.The proposed real-world 3D reconstruction method is used to reconstruct the dense 3D point cloud of the real scene.The results of the comparative analysis between the reconstructed 3D point cloud and the real scene show that the maximum error among the selected observation lines is only0.97 m,and the root mean square error is 0.68m;the multi-objective classification experiment using the Rand LA-Net algorithm for semantic segmentation of the 3D point cloud has good classification accuracy.