A Global Optimization Approach For Automatic 3D Modelling Of Indoor Structure Using Point Cloud

Human beings spend most of their time indoors.People are such familiar with the indoor environment that the 3D model of building interior structures has been awfully neglected for a long time.The slow development of techniques may contribute to that as well.Currently,2D map are widely used for indoor applications,the fire map in public buildings for instance.The demand on 3D indoor models from the Architecture,Engineering,and Construction(AEC)industry is rapidly increasing.With the advantage of direct 3D information,the data in form of point clouds has become another important source in geographic information field besides vector map and raster images.Pointclouds show a bright future in describing and expressing object,which could be applied to much more scenes.Therefore,pointclouds provide new modes for indoor modelling.Unfortunately,the pointcloud-to-Model process remains largely a manual phase due to complexity of indoor environment and the large amount,nonuniform density and quality of unstructured data.These problems block the automatic modelling of indoor structures.This paper approaches this challenge in a systemic,integrated way.The systemic analyst on indoor structures and pointclouds is first described.Then,from the multiscale views of point,voxel and planes,the key techniques of normal oritentation,semantic segmentation and modelling are deeply explored with the combination of geometry,semantic and topology.The main contents in this dissertation are sketched as follows.(1)The paper points out the characteristics of indoor environment under the depth analysis of indoor scene.These characteristics are Hierarchy,Semi-Closure,MarginRegularity,Spatial-Coherency,and Complexity.Combined with the manners of pointclouds data collection and its technical features,the paper details how pointcloud represents these five characteristics.The paper holds that the characteristics and features could guide the automatic process of indoor modelling.Then,the mathematical expression of indoor structure modelling is given.(2)Against the inconsistence of normal direction,the paper presents a global optimization method on normal direction orientation of pointclouds based on the SemiClosure characteristic of indoor scene.Furthermore,the planar voting schema in view of Margin-Regularity is adopted to refine normal direction.Experiments demonstrates that the method on normal direction orientation presented by this paper is better than the classical MST and the improved MST+QPBO-I methods.(3)To gap the absence of semantic information in indoor pointclouds,two methods on semantic segmentation are proposed according to the Hierarchy and MarginRegularity characteristics of indoor structures.The semantic segmenation is further optimized by indoor planar topology.Experiments on true data show the effectiveness and correctness of these two methods.(4)Aiming at labeling in/out(wall-slab)of indoor space,the segmentation approach on global optimization of space voxels is designed based on Semi-Closure and Spatial-Coherency characteristics of indoor structures in this dissertation.The approach could deal with complex indoor environments with Multi-storey,arbitrary wall orientations and slanting roof.It goes beyond the 2.5D and Manhattan-World assumptions.The efforts are further made to extract sub-space and storey from the voxels labeled as interior.Experiments verify the advance over the method based on height histogram.(5)It challenges the automatic modelling of indoor structure that simply relies on pointcloud data.To overcomes this problem,a global optimization framework on generating face model and solid model from pointclouds presented in this paper.Two methods on automatical modelling are evaluated on several complex synthetic and true indoor datasets,which reveals the correctness,effectiveness and advantages over current methods.