Mosaicking cluttered ground planes based on stereo vision

Abstract

Recent stereo cameras provide reliable 3D reconstructions. These are useful for selecting ground-plane points, register them and building mosaics of cluttered ground planes. In this paper we propose a 2D Iterated Closest Point (ICP) registration method, based on the distance transform, combined with a fine-tuning-registration step using directly the image data. Experiments with real data show that ICP is robust to 3D reconstruction differences due to motion and the fine tuning step minimizes the effect of the uncertainty in the 3D reconstructions.