Fast algorithm for real-time ground extraction from unorganized stereo point clouds

This paper presents a fast, robust algorithm for ground extraction from unstructured point clouds obtained from stereo reconstruction. Unlike most point cloud segmentation approaches, our algorithm does not rely on 2.5D range image structures nor on any sensor information. All processes involved consider geometry only and do not depend on any reflectivity or color information. We propose applying a top-down 4-ary segmentation followed by a segment-wise classification. This adaptive approach allows accurate differentiation between ground and obstacles in noisy point clouds of cluttered scenes. Real-time performance is achieved on a low cost embedded platform.