Building Parts Extraction And Modeling Based On Indoor SLAM Point Cloud Data

SLAM(Simultaneous Localization and Mapping)system based on threedimensional laser scanning technology is a key technology for quickly acquiring threedimensional information of indoor space.Classification based on indoor point cloud data,building component extraction and modeling are of great significance to indoor space analysis.However,the directly obtained point cloud data has the characteristics of massiveness,uneven density,severe occlusion,and noise pollution,which makes the existing indoor 3D point cloud data preprocessing methods not robust;meanwhile,the points of indoor components Cloud feature extraction makes use of large-scale spatial neighborhood information missing,point cloud segmentation methods have oversegmentation and under-segmentation,and point cloud classification network models need to be further explored.In addition,existing point cloud indoor modeling methods usually It is impossible to generate a detailed building model with complete semantics and topology,and the accuracy needs to be improved.In response to the above problems,this paper first filters the indoor point cloud data to facilitate the classification and segmentation of the point cloud data.In order to consider a wider range of neighborhood information,a recurrent neural network is used to update the characteristics of spatial points to improve the classification accuracy of point clouds,and an attention mechanism is introduced to jointly realize the semantic segmentation and instance segmentation of point cloud data,so that the two can promote each other.Finally,based on the results of classification and segmentation,the point and surface features of indoor building point cloud data are extracted and a building model containing semantics and topology is generated.The research content and work results of this article mainly include the following points:(1)Indoor room segmentation and point cloud filtering.The amount of point cloud data for indoor scenes is large.First,the point cloud data is divided into different floors according to the distribution histogram in the Z direction of the height.For each point cloud data,a triangular network is constructed on its two-dimensional projection plane and used.The graph cut method divides it into data of different rooms;for each room,an improved noise correction network is used to filter the original scan data.(2)Building point cloud classification based on feature enhanced neural network.Using segmented and denoised point cloud data as input data,construct a neural network for point cloud classification,and update the characteristics of points using a bidirectional recurrent neural network to fully consider the spatial neighborhood information;design the structure of the neural network to avoid feature degradation and gradient disappearance,So as to realize the extraction of building point cloud data.(3)Semantic and instance joint segmentation of building point cloud data.Using the metric learning method to embed building point cloud data in high-dimensional semantic space and instance space respectively,introducing attention mechanism to realize the semantic segmentation and instance segmentation of building point cloud,so that the two complement each other.(4)Three-dimensional modeling of indoor point cloud data.Use the classification and segmentation results to extract the point cloud data of indoor point cloud data such as ceiling,floor,wall,door,window and other fixed structures,and then use the segmentation and boundary optimization algorithms to extract the key points and surface information of the indoor model and Its semantic information,and verify the results based on geometric and semantic information,and finally generate an indoor model with complete semantic and topology.In this paper,the publicly available S3 DIS dataset,ScanNet dataset and measured Nanjing Normal University indoor dataset are used to verify the proposed algorithm.The experimental results show that the feature point-based neural network-based building point cloud classification algorithm proposed in this paper improves the classification accuracy of building point cloud data,and the joint semantic and instance segmentation algorithm improves the classification and segmentation accuracy of building point cloud data;The three-dimensional modeling method of indoor point cloud data can generate geometric and semantically rich indoor building models,which improves the automation of indoor modeling.