Surface-based Segmentation of 3D Range Data

This article presents a holistic approach to the surface-based segmentation of 3D range data. The proposed framework is generic and requires only that i) the data be dense enough to allow for normal vector estimation and ii) the view point be known for every point in the data set. In contrast to many previous works on segmentation that consider only a few unoccluded and densely sampled objects, we describe and demonstrate a segmentation toolchain that is applicable to sparse and noisy point clouds as e.g. obtained by a mobile robot with an actuated laser range finder. The data may consist of several registered views, may contain partial occlusions and may exhibit variable density. The scenarios used for evaluation consist of simulated as well as real point clouds acquired in different human-friendly environments. The results show that the proposed framework yields a meaningful segmented geometric abstraction of the examined scenarios with segmentation times on the order of the scan times. It is therefore of high relevance to real-life robotics scenarios, where segmented surfaces can e.g. serve as input for object recognition algorithms. The tradeoffs and limitations of the approach are outlined and further potential use is discussed.