Efficient segmentation and plane modeling of point-cloud for structured environment by normal clustering and tensor voting

In this paper, we introduce an efficient point-cloud segmentation algorithm, inspired by efficient segmentation (also named as super-pixel extraction). It uses parameterised “normal words” as distance measures, which are obtained by clustering of surface normals. We estimate the surface normals by the sparse tensor voting framework, which enables adaptive structural extraction, even for the case of missing points. The output result is consist of labeled point representations regarding plane assumptions, which is validated by metrics based on information theory. We show the quality of the segmentation results by experiments on real datasets, and demonstrate its potentials in aiding 2.5D topological navigation for structured environments.