Research On Mesh Refinement And Texture Mapping Based On Oblique Multi-view Images For 3D Urban Model

The process of real 3D model reconstruction can be divided into two parts:geometric reconstruction and texture mapping.An accurate geometric model can reflect the real shape of the model,and a rich texture can show the real appearance of the model.With the rapid development of photogrammetry,the efficiency of data acquisition is increasing,and the speed and accuracy of data processing are also improving.In the field of 3D model reconstruction,a lot of research focuses on the restoration of 3D model geometry.Now,a set of mature theoretical system of 3D model reconstruction has been formed,while the research on texture mapping of large-scale 3D model is relatively less and has no significant breakthrough.As the object of texture mapping,the quality of geometric model largely determines the quality of texture.In the complex urban environment,the existing geometric model reconstruction results are not enough to describe the details of the model.In this paper,the research is divided into two parts:mesh refinement research and texture mapping research.Firstly,a mesh refinement method based on photo consistency is introduced.And on this basis,how to ensure a large number of lines in the scene,such as wall lines and stair lines,is further discussed.Then,after a brief introduction of each technical process of texture mapping using oblique multi-view images,texture misalignment and low utilization of texture layout occurred in texture mapping are deeply studied,and a texture mapping scheme with higher accuracy and higher realism is proposed.The specific research contents and works are as follows:In the research of mesh refinement,this paper describes how to use the rich image information to refine the mesh.The process is to first build the energy function on the surface,which is composed of photo-consistency and regularization.Using the reprojection error to measure the photo consistency of the model projected on different images,and Laplace smoothing forms the regularization term.Finally,minimizing the energy function by variational refinement method.On this basis,this paper analyzes the problem that this method can not well express the characteristics of the lines in the 3D urban scene environment,and proposes a mesh refinement method with line constraints,which effectively expresses the straight and sharp characteristics of the line in the 3D urban scene.In the research of texture mapping,this paper first systematically introduces the texture mapping technology based on oblique multi-view images,including visibility analysis,view selection,color adjustment and texture layout.Then,texture misalignment at the seam and low utilization of texture layout are discussed respectively.The causes of texture misalignment is that the model and image can not be completely aligned due to the inaccuracy of camera poses or the inaccuracy of reconstructed model in the process of texture mapping.In this paper,a method is proposed to guide the view selection of Markov Random Field by using the projection error obtained from SFM(structure from motion).The image with better geometric quality is selected as much as possible for each triangle,and then texture misalignment is further corrected with image deformation.By designing the fusion criteria,the small texture patches produced by the view selection are fused into the large texture patches to further reduce the generation of seams is studied.Finally,a texture layout method based on greedy algorithm is proposed,which can effectively improve the layout utilization of texture atlas by searching the remaining space.