Laser Point Cloud 3D Real Modeling Based On Fused Sequence Images

3D laser scanning technology has been rapidly developed in the field of scene digitization,and a large number of real scenes require fine-grained 3D models.Because the 3D laser scanning technology can only meet the restoration of geometric position,but it can not achieve satisfactory visual effect in texture.In order to construct the 3D model of the scene with high precision,this paper combines two kinds of heterogeneous data,the terrestrial laser point cloud and the sequence images in three-dimensional space,which avoids the difficulty of cross-dimensional data processing.The two feature registration strategies are combined with significant feature extraction and geometric relationship for point cloud registration.Based on the efficient search principle of bidirectional Kd tree,the texture mapping is used to repair the holes in the point cloud,and finally a complete 3D model is generated.The main research contents of this paper are as follows:(1)Fully consider the topographical factors and the structural features of the scanned objects in the scene,and combine the grid height limit,the normal angle,and other judgment criteria to separate the ground points from the non-ground points,use the clustering idea that the samples of the same category are closely connected,scanned objects at different locations are extracted.(2)On the basis of the two registration strategies corresponding to the feature description and the geometric relationship for point cloud registration,the ISS feature points are evenly distributed but the number is small,and the defects that effectively describe the regions with large curvature changes are introduced.The significant constrained conditions were introduced to extract additional salient points on the surface of the point cloud,and the extracted features were analyzed by multi-scale.The stable feature points were preserved under three different search scales,and the initial registration was performed in combination with the Super4 PCS algorithm.On this basis,the results of the initial registration are refined by using the Sparse ICP,and the multisite cloud is superimposed to generate a complete point cloud model.(3)In order to transform the image data into three-dimensional space with high integrity and high texture precision,compare and analyze the three-dimensional reconstruction principle and flow of VisualSFM,COLMAP and OpenMVG.After 3D reconstruction,the color of the model surface is calculated in RGB and HSV space with the true image color as the true value.The color distance between the pixels and the color similarity between the images are evaluated.Based on the principle of Kd tree neighborhood search,the mapping relationship between the image point cloud with high texture precision and the laser point cloud with high geometric position accuracy is established.At the same time,the hole boundary of the extracted laser point cloud is inversely mapped to the image point cloud based on Kd tree.According to the established parallel equidistance principle,the holes in the laser point cloud are filled with dots,and the texture mapping of the laser point cloud surface is perfected.After the information fusion of the point cloud model(including closed data and non-closed data),by comparing the flow and advantages and disadvantages of greedy projection triangulation,Poisson surface reconstruction and marching cube,choose greedy projection triangulation to construct the surface,and the complete 3D surface model is generated after streamlining and smoothing.