Robust RGB-D Camera Motion Estimation using ICP-based Edge Alignment

Abstract

This paper presents a robust RGB-D visual odometry (VO) using iterative closest points (ICP)-based image edge alignment. Most existing VO algorithms are mainly depending on two assumptions about operating environments

brightness consistency and texture-abundant scenes. However, in real situations, the brightness is non-uniformly changing and low-textured scenes, such as monotonic wall and ceiling, are often encountered. In these cases, general VO could yield degraded performance or, at worst, lose track of estimations. Our primary concentration is to make VO more stable and robust against to these challenging environments. To achieve this, we estimate camera motions using image edge pixels which are prominently observed even in low-textured scenes with the changing brightness. For more stable ICP optimization procedure, we propose robust edge matching criteria exploiting image gradient vectors. Through extensive experiments, we investigate the robust performance of proposed edge pixel matching criteria and its stabilizing effect on overall optimization steps. We evaluate estimation performances of the proposed method using RGB-D datasets with various brightness changing conditions and less textured scenes. Our approach shows robust and competitive performances against to challenging scenes with less textures and varying brightness, where several state-of-the-art methods are considerably degraded and sometimes fail to estimate the camera motions.