Study On The Parallel Implemention And Application Of 3D Reconstruction Algorithm With GPU Cluster

The rapid development of unmanned aerial vehicle(UAV)technology and the continuous expansion of its application field makes it a widely used new remote sensing data acquisition platform.Compared with aerial photography and satellite remote sensing platforms,UAV remote sensing platform has the advantages of low cost,simple operation and fast data acquisition.The UAV 3D reconstruction technology can obtain the texture information of the side of the ground object through multi-angle and multi-directional shooting,which can realize the large-scale,high-reality and high-precision 3D city modeling,greatly improving the operability of simulated 3D modeling,and get extensive attention and applications.Compared with traditional modeling methods,using UAV for 3D real-world modeling has a significant improvement in data acquisition and processing efficiency.However,when using these commercial softwares,there will be the following problems:(1)Compared with the traditional manual modeling method,using commercial software for 3D modeling has greatly improved the efficiency,but the processing cycle is still long when large-scale operations are carried out.(2)The hardware environment requirements of commercial software are constantly increasing,and ordinary desktops often fail to meet the performance requirements,which also brings a lot of inconvenience to a certain extent.(3)Although commercial software can also support parallel processing in a local area network,communication deadlocks and missing tasks often occur,and commercial software is mostly closed source and cannot be modified and developed.In order to solve these problems,this research is based on GPU cluster and parallel computing technologies,and studies a complete,open and fast parallel 3D reconstruction process and the key algorithms.The main research contents include:(1)Research on 3D reconstruction optimization based on MPI parallel computing and CMVS computer vision algorithm.Based on the computer vision related algorithms,build an open source,optimized 3D model reconstruction process chain;According to the hot spot analysis results of the serial 3D reconstruction process,locate the performance bottlenecks of the process;Based on the in-depth analysis of the algorithms' principle and process,analyze and determine the parallelism and parallel optimization method of the algorithms,and realize the CMVS algorithm based on the MPI parallel programming mode.(2)Research on parallelism optimization of key algorithms of 3D reconstruction process under GPU environment.In order to further speed up the 3D reconstruction process and take advantage of GPU in processing graphics and images,this study analyzes the 3D reconstruction process in detail.Utilizing the characteristics of GPU hardware,the parallel algorithm of dense reconstruction is designed with the hot spot analysis results.And then MPI+GPU two-levels parallelism of the 3D reconstruction method in the GPU cluster environment in combination with the research content(1)is acchived.In addition,GPU parallel technology is also used to process feature extraction of dense reconstruction,and further optimized by means of combining kernel functions.Finally,in the sparse reconstruction process,the relevant processes are improved by referring to the GPU optimization method in the dense reconstruction process,so as to realize accelerated sparse reconstruction processing in the GPU environment.(3)Relevant experiments are designed to verify the effectiveness of the optimization method proposed in this paper.After the main optimization process of this paper is completed,the 3D dense point cloud data set is taken as an application example,and the relevant point cloud post-processing technology is used to realize 3D point cloud surface reconstruction and generate 3D models.