Research On Fast 3D Reconstruction Technology Of Field Scene Based On UAV Image

With the continuous development of modern information technology,people have higher and higher requirements for the clarity of information representation in the physical world.However,the three-dimensional(3D)model can describe a lot of intuitive information of objects,which is widely used in scene construction.The 3D reconstruction of the outdoor environment,especially the efficient and robust 3D reconstruction,has attracted more and more attention in the civil and military fields.For the 3D reconstruction of large field scenes,traditional methods mainly include manual measurement,satellite remote sensing mapping,and lidar scanning and reconstruction.However,these traditional reconstruction methods have some disadvantages such as high cost,long time-consuming,demanding and so on.In recent years,consumer unmanned aerial vehicles(UAV)have been developing rapidly.They are widely used in field aerial photography scenes due to their characteristics of convenient carrying,flexible deployment and wide coverage area.The rapid 3D reconstruction of the field scene based on UAV image sequences has broad application prospects.However,there are still problems such as high algorithmic complexity,and poor timeliness,which are difficult to meet the application needs of the industry.To solve these problems,this subject mainly studies a fast three-dimensional reconstruction method of field scene based on UAV image.Through the study of computer vision,motion recovery structure and multi-view reconstruction,combined with the background of field maneuvering applications,the main purpose is to improve the reconstruction speed,and control the accuracy within a certain range.Realize a complete set of rapid 3D reconstruction process for large-scale outdoor scenes,and carry out the landing in combination with the project.The main work and results completed in this thesis are as follows:(1)A parallel processing algorithm for UAV images is proposed,which is based on the two computing architectures of CPU and GPU to accelerate the parallel processing in the two stages of 3D reconstruction,so as to improve the execution speed of 3D reconstruction task.First,the CPU-based parallel processing strategy adopts the Open MP multithreading mechanism for multi-core processors to load UAV images in parallel,so as to provide efficient data sources for subsequent processing.The second is the GPUbased parallel processing strategy.By improving the SIFTGPU algorithm,the image features are extracted in parallel on GPU to provide feature input for fast reconstruction.In the image processing stage,the algorithm proposed in this thesis can realize automatic scale optimization of UAV images and instantaneously extract GPS information,to obtain the adjacency relationship between UAV images.In the feature extraction stage,the number of feature points can be greatly increased under the premise of ensuring quality.(2)A segmented structure from motion algorithm for UAV images is proposed.According to the GPS information in the UAV images,the spatial position of the images is sorted according to the longitude and latitude.The UAV images are divided into multiple image groups.There are the same UAV images between the adjacent image groups.The SIFTGPU algorithm is used to extract the features of the UAV images.The global SFM method is used to reconstruct the sparse point cloud of the scene in parallel.Finally,according to the grouping order of UAV images,the reconstruction results of each image group are fused.And all the 3D space points and the poses of camera are optimized.Through image grouping,the accumulation of error can be reduced and the drift can be eliminated.On the premise of ensuring the accuracy of 3D reconstruction,the speed of 3D reconstruction is improved.(3)A new mesh blocking algorithm for texture mapping is proposed.Design a mesh blocking algorithm to cut a large mesh data into many small blocks.Determine the size of the available memory based on the memory usage of the compute unit.Establish the corresponding relationship between the size of memory and the size of image block.According to the size of the image block,the size of the corresponding mesh block in the mesh local coordinate system is calculated,and then the size of the mesh block is determined.According to the performance of the computing processing unit,the size of mesh block is determined adaptively.The memory pressure is reduced through block division,reducing the requirement for computing resources,and improving the speed of3 D reconstruction.(4)A rapid 3D reconstruction prototype system for field real scene is built.In the way of modular development,a rapid 3D reconstruction prototype system for field real scene is built.It mainly includes: UAV aerial photography module,sparse point cloud construction module and 3D scene generation module.The prototype system is decoupled through modular processing.In the calculation process,each module does not affect each other,so as to solve the problem of complex calculation in the prototype system.The specific steps include setting the target area for UAV in the field real scene,and taking aerial photography through the four-rotor UAV according to the set flight path.According to the images of UAV,the 3D reconstruction of the field real scene is carried out to verify the function and performance of the prototype system.In summary,this subject studies a fast 3D reconstruction method for field scenes based on UAV images.It is mainly aimed at 3D reconstruction of large field scenes.Lightweight algorithms are designed for the key links of reconstruction to reduce the consumption of computing resources.This thesis solves the problems of slow speed,poor timeliness and poor robustness of 3D reconstruction for large field scenes.In this thesis,the prototype system is designed and built,and the methods in this thesis are connected in series in the prototype system.The UAV images are used to reconstruct the field scene,and then the efficiency of the methods in this thesis are verified.