Point Cloud Object Segmentation Using Multi Elevation-Layer 2D Bounding-Boxes

Tristan Brodeur, Hadi AliAkbarpour, Steve Suddarth; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2021, pp. 3917-3925

Segmentation of point clouds is a necessary pre-processing technique when object discrimination is needed for scene understanding. In this paper, we propose a segmentation technique utilizing 2D bounding-box data obtained via the orthographic projection of 3D points onto a plane at multiple elevation layers. Connected components is utilized to obtain bounding-box data, and a consistency metric between bounding-boxes at various elevation layers helps determine the classification of the bounding-box to an object of the scene. The merging of point data within each 2D bounding-box results in an object-segmented point cloud. Our method conducts segmentation using only the topological information of the point data within a dataset, requiring no extra computation of normals, creation of an octree or k-d tree, nor a dependency on RGB or intensity data associated with a point. Initial experiments are run on a set of point cloud datasets obtained via photogrammetric means, as well as some open-source, LIDAR-generated point clouds, showing the method to be capture agnostic. Results demonstrate the efficacy of this method in obtaining a distinct set of objects contained within a point cloud.