SfM Algorithm Of 3D Reconstruction From UAV Aerial Images

3D reconstruction based on images has always been a hotspot in the field of computer vision.As a key technology and process of it,sparse points cloud reconstructed by SfM(Structure from Motion)directly affects the quality of the final reconstruction model.For another hotspot SLAM(Simultaneous Localization and Mapping),the spares points cloud also determines the accuracy of environmental mapping and pose estimation of vehicle.SfM algorithm for large-scale scenes based on UAV(Unmanned Aerial Vehicle)aerial images are studied in this paper.Analyzing the basic principle of SfM,the major SfM algorithms based on epipolar geometry are divided into incremental,global and hybrid types according to the algorithm differences respectively.Regarding the robustness,completeness and accuracy of as indexes,the 3 algorithms are tested and compared by the public datasets.On the basis,an improved SfM algorithm is proposed.Since the global SfM has good performance when the correlation among images is good,and the incremental SfM has strong anti-interference ability when noise exists,the combination of the 2 is helpfule to improve the quality of reconstruction.Moreover,the process of clustering,segmentation and noise filtering added to the global SfM can further improve the performance.The details of rotation averaging,hierarchical clustering,global rotation estimation,major clusters rebuilt and camera pose estimation are discussed.Experiments of proposed algorithm and three major ones are carried out by public datasets and self-collected UAV aerial images,which verify the effectiveness of the improved algorithm.Considering the rapid development of deep learning in the field of computer vision in recent years,3D reconstruction based on deep learning may be a new development trend in this field,which also be explored in this paper.The principle of SfM based on deep learning(DSFM)is analyzed,and 3D reconstruction experiments are carried out by DeMon,an open source software package.The accuracy of DSFM and traditional two-view SfM is tested and compared.The 2 types are also be tested for public datasets and self-collected aerial images.