Supervoxel-based Graph Clustering for Accurate Object Segmentation of Indoor Point Clouds

The correct segmentation of clustered objects in indoor scenes is a critical step of machine understanding in the field of robotics. However, traditional segmentation methods can hardly meet the requirements of accuracy and real time. To this end, an unsupervised object segmentation method based on supervoxel and graph clustering is proposed in this paper. First, the background is removed based on RANSAC, and the remaining point cloud is over-segmented into supervoxels, which is used to improve the segmentation efficiency and to obtain fully connected graph. Second, a graph clustering method based on pure geometric features is proposed to obtain the corresponding subgraphs of each object using the concavity between supervoxels. Finally, a new refinement method for processing the boundaries at the intersection is adapted for obtaining the boundaries with point-level accuracy to improve the segmentation accuracy. The method is experimented on both researcher community public and self-collected datasets. Qualitative and quantitative results show the applicability of our method to robotic real-time object manipulation and its broad development prospect.