Abstract: Indoor scene point cloud segmentation is the basis for the perception,understanding,analysis and application of 3D scenes. Although the current indoor point cloud segmentation method can be applied to many scenes,it lacks the ability to adapt to the segmentation of different structures,and the bottleneck of undersegmentation cannot be avoided when dealing with segmentation of near parallel surfaces. Therefore,a method based on improved supervoxel and graph cut is proposed,through the neighborhood selection,normal vector calculation,feature distance measurement,and spatial connectivity constraint and segmentation of supervoxel local adjacency graph,an adaptive mechanism for processing the relationship between adjacent parallel surfaces is constructed,which realizes the effective segmentation of point cloud in complex indoor scenes,and the above-mentioned undersegmentation phenomenon is overcome. Finally,the proposed method is verified by four indoor scene point clouds,and compared with the existing methods. Experimental results show that the proposed method improves the accuracy and recall of point cloud segmentation in complex indoor scenes,and verifies the adaptability and effectiveness of the method to different structural segmentation.

Key words: indoor scene, point cloud segmentation, supervoxel, graph cut