Research On Supervoxel Based Region Growing Segmentation For Point Cloud Data

Point cloud segmentation is the significant part in point cloud processing,and is the fundament of 3D reconstruction,scene understanding and target recognition.It is currently one of the most important research subjects in the point cloud segmentation field to optimize the design of the segmentation algorithms and improve the accuracy of segmentation results.On the basis of summarizing the research status of exsting point cloud segmentation methods,for the issue that existing segmentation methods result in rough boundary and unsmooth region on the segmentation result,our paper proposes a region growing segmentation method based on supervoxel for point cloud,which can avoid under-segmentation and oversegmentation effectively and realize precise segmentation of point cloud.The main works achieved in this paper are as follows:Over segmentation to get supervoxels.Over segmentation preprocessing is implemented on point cloud to obtain smooth supervoxels attached to object boundary.Firstly,in the threedimensional space,the unstructured point cloud data are transformed into voxel cloud with fixed resolution,and the adjacency graph of voxels is established according to 26 neighborhoods in the voxel space by traversing the KD tree.Secendly,filter seed voxels by meshing to initialize the clustering method with flow constraint.Thirdly,on the basis of the adjacency graph,the distance of the neighborhood voxels is measured by considering the spatial position,geometric features and boundary information.Finally,according to the flow constraint clustering method,cluster the voxels to get supervoxels and realize the over segmentation of the voxel cloud.Cluster supervoxels based on region growing.Fuse supervoxels by using the principle of region growing to achieve the segmentation of point cloud data.Firstly,the method performs plane fitting on the supervoxels,and uses the residual value of plane fitting to measure the curvature change of the surface.A reasonable residual threshold is used to initialize the supervoxel seeds,which ensures that the segmentation method can distinguish thoes points with similar normal but belonging to different objects.Secondly,the method introduces the Point Feature Histogram to capture the geometrical features of surface and use the variety of the normal direction to ensure the smoothness of segments.Our method merge supervoxels by considering geometric features and smoothness constraints,avoiding the problem of over-segmentation and under-segmentation.Finally,the similarity between adjacent supervoxels is calculated by normalizing the eigenvector,and the local fusion of supervoxel is realized by the principle of region growing.In this paper,we propose a region growing segmentation method based on supervoxel,which use the conception of “first over segmentation,and then clustering” to realize the exact segmentation of point clouds.Comparative experiments are performed for several sets of different scenes.The results show that,the segmentation method proposed in this paper not only preserves the simplicity and high efficiency of original region growing method,but also reduces the influence of noise and outliers to the segmentation results and enhances the stability and accuracy to deal with complex scenes.While improving the efficiency of the segmentation,this method ensures that the final results have smooth segmentation regions and precise segmentation boundaries.