Research On Modeling Method Of Typical Indoor Object Based On Model Library

With the development of 3D reconstruction technology of outdoor scenes,the demand of 3D reconstruction of indoor scenes is increasingly prominent.The construction of indoor 3d scenes provides important technical support for virtual reality,augmented reality,indoor positioning and navigation,emergency response,public security and other fields.At present,the commonly used data sources for indoor scene reconstruction can be divided into images and laser point cloud data.The results of interior reconstruction based on image data are greatly affected by the conditions of interior texture and illumination,and the modeling takes a long time.The 3D point cloud data of indoor scene can be acquired directly and quickly by using laser scanner to collect the cloud data of indoor scene for the reconstruction of indoor scene,which is an efficient data acquisition method.However,due to the lack of color and semantic information of the laser point cloud data,it also brings a lot of inconvenience to the actual use process.However,structural modeling on point cloud data can effectively solve the above problems.This modeling method can quickly obtain the key structure and topology information of indoor scenes,and provide important technical support for the emergency and other fields.The reconstruction of indoor objects based on model library is an essential technique in the process of structured modeling.The main research contents of this paper are as follows:1.The coarse registration method of indoor scene cloud was studied.A method of point cloud registration based on the edge and corner information o f point cloud projection contour by indoor laser scanning is proposed in this paper.In view of the possible mismatches caused by the high symmetry of indoor scene,the adjacent topological graph is constructed by using the edges and corners,and a certain constraint is carried out by the topological relation.The possible error registration results of the scene were eliminated and the final registration results were verified by the verification method of maximum consensus set.2.The semantic segmentation and optimization methods of indoor scanning point cloud are studied.On the basis of semantic segmentation of indoor point cloud by using deep network,in view of the rough boundary and label errors in some areas in the initial segmentation results,the full-connection condition random field method is used to optimize the segmentation results,so as to improve the segmentation results,and the labels of each point have a higher confidence.3.A 3D model library of indoor objects is established and a model-driven modeling method for typical indoor objects is studied.A typical indoor object modeling method with translational rotation and scale invariance is proposed to solve the problem of automatic retrieval and matching of indoor object point cloud and 3D model library.To solve the problem that it is difficult to calculate the degree of similarity between point cloud data,point cloud data is converted into voxel data for calculation.