Research On 3D Reconstruction Technology Based On UAV Mapping

With the development of oblique photography technology and artificial intelligence,3D reconstruction technology of UAV mapping images has made great progress in recent years.Taking outdoor large-scale scenes as the main research object,the method of restoring their three-dimensional structure with the help of UAVs is efficient and accurate,and has been successfully applied in urban planning,remote sensing,mapping and other fields.3D reconstruction of UAV mapping requires taking images from different angles and lighting conditions.For large outdoor scenes,there are problems such as large target scale,obstruction,bad weather,which make it difficult to reconstruct efficiently.The accuracy of reconstruction results is also easily affected by weather,wind speed,illumination,camera shake and other factors,so a robust reconstruction algorithm is needed to meet the needs of high precision.In addition,the current common UAV image reconstruction algorithm takes a long time,and there is still a large space for improvement in accuracy.Therefore,this paper takes outdoor large-scale scene as the application scene to study the 3D reconstruction algorithm of UAV images.The main contents are as follows:1.An explanation of the theory of multi-view geometry.The model and calibration principle of the pinhole camera are described,and the mathematical principle is deduced.Three common feature extraction descriptors and feature matching algorithms are introduced.The advantages and disadvantages of various SFM methods are analyzed and discussed.The basic principle of dense reconstruction algorithm based on multi-view stereo is expounded and deduced.The incremental SFM was used to reconstruct and analyze the results.2.The learning-based multi-view stereo algorithm is studied,and an improved PointMVSNet neural network is proposed.The self-attention mechanism is introduced,the attention principle and mathematical derivation of scale agnostic attention are given.A scale-agnostic feature pyramid attention module was built to capture the long-range feature correspondence,and the local receptive field was expanded to enhance the feature.The accuracy and completeness of the improvement network are verified on the dataset,and the visualization results of large-scale outdoor scenes mapped by UAV are presented.3.By analyzing the advantages and disadvantages of reconstruction algorithms based on SFM and deep learning,a 3D reconstruction system of self-collected data based on deep learning is proposed to realize robust and fast 3D reconstruction of arbitrary self-collected data.Through this system,four kinds of outdoor scenes were verified experimentally,and the influence of image data on the results was analyzed.4.The 3D reconstruction system based on deep learning proposed in this paper is applied to engineering practice.The UAV is used as the experimental platform to carry out real outdoor scene sampling and reconstruction experiments,and the input is expanded through preprocessing.Experimental results show that this method has good accuracy and robustness,and is convenient and efficient in application.It has certain engineering application value.In addition,the real-time 3D reconstruction based on UAV mapping was studied,and the experimental analysis was carried out with ORB-SLAM2.