Indoor Multi-source Point Cloud Planar Segmentation Method Based On Normal Saliency

The continuous promotion of smart city construction has led to an increasing demand for fine perception and 3D modeling of urban building interiors.Applications such as indoor navigation,building completion and acceptance,building maintenance and renovation planning have raised the need for rapid 3D reconstruction of the main spatial structures of indoor scenes,while indoor point cloud segmentation is one of the important steps in indoor3 D modeling.Indoor scenes are characterized by complex structures,numerous objects,and occluded areas,etc,meanwhile,with the continuous development of 3D laser scanning technology,lightweight and flexible low-cost sensors have been widely used to obtain point cloud data of indoor scenes,and the data quality of indoor scenes point cloud data has become more and more uneven,which have caused some difficulties to the indoor scenes point cloud segmentation.The existing point cloud planar segmentation methods usually rely heavily on the quality of point cloud data,which makes it difficult to segment complex indoor scene point clouds accurately and efficiently,and cannot meet the demand for fast segmentation of indoor scenes and obtaining information about the main structures and objects in the house(within the floor).To address the above problems,this paper proposes a method for segmenting indoor multi-source point cloud planes based on normal saliency,which can effectively segment planes in indoor and corridor scene point clouds.The main work contents of this paper are listed as follow:(1)An indoor point cloud planar segmentation method based on normal direction saliency is proposed.Based on the characteristics of indoor scene point clouds,this paper proposes an indoor point cloud planar segmentation method based on normal direction saliency.Specifically,the method first pre-segments the indoor scene point cloud using super voxels and establishes topological relationships between neighboring super voxels;then projects the super voxel normal vectors onto the surface of the Gaussian hemisphere according to their corresponding directions,and achieves a fast planar grouping using the K-means method;finally,instantiates the planar segmentation according to the topology of the super voxels.(2)A global energy optimization method for planar segmentation with multiple primitives interacting with each other is proposed.In order to improve the accuracy and robustness of planar segmentation,this paper proposes a planar segmentation global energy optimization method with multi-level interactions among primitives.The method suggests releasing points from the outlier super voxels containing corner points and plane edge points and develops multi-level topological relationships with three primitives from different stages;then a global energy optimization strategy is introduced to establish a third-order energy equation based on the interactions among multi-level primitives to improve the over-segmentation problem.In order to verify the feasibility and effectiveness of the proposed method,the indoor modeling benchmark data set provided by ISPRS and self-collected multi-source point cloud data were used to test the proposed method.By comparing with three other advanced point cloud plane segmentation methods,it is found that this method can obtain reliable plane segmentation results and has the optimal plane segmentation performance.The experimental results prove that the method in this paper can achieve the effective segmentation of cloud planes of indoor field points in a complete,efficient and accurate way,while adapting well to the indoor multi-source point cloud data collected by different sensors.