Research On The Generation Of Image Dense Matching Point Cloud Based On Incremental SFM And Their Fusion With Laser Point Cloud

With the advancement of science and technology,and the rapid development of sensor technology,3D laser scanning technology is gradually transitioning from‘professional-grade’ to ‘consumer-grade’ and has been widely used in information construction such as urban 3D modeling and preservation of cultural heritage.3D laser scanning technology can quickly and accurately obtain massive surface information of target objects without contact,and has low susceptibility to external interference,which has attracted widespread attention from researchers in the field of surveying and mapping.However,3D laser scanning technology lacks texture and color information when capturing surface information of target objects,and when applied in urban areas,the acquired point clouds often contain sampling gaps due to occlusions,instrument field of view,and other factors.Currently,interpolation methods are commonly used to fill the sampling gaps in Li DAR point clouds.However,when applied to the surface reconstruction of complex urban buildings,these methods may not accurately restore the surface features of the buildings.Therefore,there is an urgent need to find a reliable approach to fill the sampling gaps in Li DAR point clouds.Images,on the other hand,are easy to acquire and contain rich texture and spectral information.By combining Li DAR point clouds with images,it is possible to directly obtain the spatial coordinates and texture information of objects,and fill the gaps in Li DAR point clouds.This approach has wide potential applications in urban 3D modeling,object classification,and other fields.Based on the above analysis,this thesis aims to achieve the complementary advantages of Li DAR point clouds and images.Using Li DAR point clouds and images as data sources,the high-precision fusion of Li DAR point clouds and images is chosen as the starting point for the research.The specific research work conducted is as follows:1)Sparse point clouds of the target objects are generated by using Structure From Motion(SFM)approach.To address the issues of time-consuming and unstable model that may result in reconstruction failure in current SFM methods,a dual-selection addition strategy-based incremental SFM method is proposed.This method improves the efficiency of the algorithm while enhancing its robustness in generating sparse point clouds of the target objects.Based on the sparse point clouds obtained from images,dense point clouds are generated using CMVS-PMVS for image dense matching.2)Registration and fusion of laser Li DAR point clouds and image dense matching point clouds.To address the problem of scale inconsistency between laser Li DAR point clouds and image dense matching point clouds,the target scale of the 3D laser point cloud is used to achieve data scale consistency through directed bounding boxes.Based on Fast Point Feature Histograms(FPFH)and 4 Point Congruent Sets(4PCS),the coarse registration of laser Li DAR point clouds and image dense matching point clouds is achieved by selecting feature points that represent point cloud information from FPFH using principal component analysis and using them as original data for 4PCS registration.This reduces the size of the search point set,improves computational efficiency,and solves the problem of unreasonable downsampling in 4PCS algorithm.Then,the ICP algorithm is used to achieve precise registration of laser Li DAR point clouds and image dense matching point clouds.Finally,on the basis of satisfactory registration accuracy,the correspondence between Li DAR point clouds and image dense matching point clouds is established through kd-tree search,and texture mapping is performed to achieve fusion of the two datasets.By using the RIEGL VZ-1000 3D laser scanner(equipped with a Nikon D600camera),point cloud and image data of the Jiaozuo Road Mining School plaque at the Nanhu Campus of China University of Mining and Technology were collected.The collected laser point cloud data and image data were fused based on previous research results,and a complete 3D surface model was obtained through Poisson surface reconstruction using Mesh Lab software.This 3D surface model can be used for visualization,analysis,and other applications,providing an accurate 3D model for the research and application of the Jiaozuo Road Mining School plaque.There are 50 figures,8 tables and 67 references in this thesis.