In this paper, a novel method for a multi-layer description of the environment based on curvature information is presented. The proposal consists of three consecutive stages: data acquisition and pre-processing, segmentation and landmarks extraction. The main novelty of this work is the use of an RGBD sensor that projects the 3D points onto a set of planes to different heights. Each one of these planes corresponds to different scan laser readings that are segmented using an adaptive curvature estimation. This adaptive segmentation is used to directly extract breakpoints,line segments and real and virtual corners of the 3D environment. Experimental results show that the proposed approach is efficient for detecting landmarks in structured and semi-structured real and virtual environments. A comparative study of similar approaches in terms of robustness and accuracy demonstrates the improvements of the presented environment description system.
|Title of host publication
|Proceedings of Workshop of Physical Agents
|Published - 2014