Research On Indoor Point Cloud Segmentation Algorithm Based On Convex Fusion Of Adjacent Regions

In order to achieve accurate segmentation of objects in a point cloud scene,the point cloud segmentation is devoted to designing an efficient and robust segmentation algorithm.It is a key basic step and important research topic in vision application fields such as target detection,three dimensional reconstruction,scene understanding,and object recognition.Aiming at the problems of low efficiency,rough boundary of object segmentation and incomplete segmentation of the existing point cloud segmentation methods,this article presents an indoor point cloud segmentation algorithm based on convex fusion of adjacent regions in the face of 3D point cloud data of indoor scenes,and achieves efficient semantic segmentation of objects in point cloud scenes.The main work and innovations of this article are summarized as follows:A point cloud supervoxel over-segmentation method based on boundary constraints is proposed.Aiming at the problems of supervoxel generated by the existing over-segmentation method crossing the boundary of the object,by extracting the boundary information of the point cloud data used to constrain the generation process of the supervoxel,a series of supervoxels strictly attached to the object boundary is obtained,improves the accuracy of subsequent segmentation.Firstly,voxelization is performed on the point cloud data.Based on the variation of the surface normal vector,the boundary of the object is extracted using the Canny edge detection method,and the boundary voxel is determined.Subsequently,the seed voxel is selected,and the spatial distance and the geometric characteristics are comprehensively considered to perform voxel similarity measure,using flow-constrained clustering algorithm,iteratively voxel clustering to generate supervoxels from seed voxels,and using boundary information for clustering constraints.For the edge point voxel that have been expanded into supervoxels,no more neighbor expansion is performed,so that the supervoxel generation process stops when it encounters the boundary,and the supervoxel that strictly adhere to the object boundary are obtained.A region fusion method based on the concavity and convexity of adjacent patchs is proposed.A series of segmented patches is obtained by using region growing for the supervoxels obtained by over-segmentation.The concavity and convexity of adjacent patches are calculated,and the convexly connected patches are merged to improve the completeness of the object segmentation.This achieves the accurate semantic segmentation of objects in point cloud scene.Firstly,the normal and the curvature of a supervoxel is calculated by the least squares fitting method.The supervoxel with the smallest curvature is selected as the seed point,and the supervoxel normal and point feature histogram are integrated to perform supervoxel similarity measure.The generation of the scene segmentation planes by using the principle of region growing.Subsequently,by analyzing the structural characteristics of the objects in the indoor scene,according to the included angle relationship between the normal vectors of adjacent patches,the concavity and convexity of the adjacent patches of objects are defined,and the convexly connected patches are merged into one region,making the final segmentation result correspond to the entire object in the scene.In this article,we propose an indoor point cloud segmentation algorithm based on convex fusion of adjacent regions,which use the idea of “first over segmentation and then fusion” to implement the semantic segmentation of objects in the indoor scene of point cloud.the supervoxel generation process use the constraint of the boundary and the region fusion method based on the concavity and convexity of adjacent patchs are innovative.The experimental results show that the supervoxels extracted by this method are uniform in size and strictly adhere to the boundary of the object,which is better than the VCCS algorithm.The final segmentation result has a clear boundary,and the segmented object is complete and accurate,avoiding over-segmentation and under-segmentation issues,effectively improves the accuracy of scene semantic segmentation.