Evaluation of recent approaches to visual odometry from RGB-D images

Abstract

Estimation of camera motion from RGB-D images has been an active research topic in recent years. Several RGB-D visual odometry systems were reported in literature and released under open-source licenses. The objective of this contribution is to evaluate the recently published approaches to motion estimation. A publicly available dataset of RGB-D sequences with precise ground truth data is applied and results are compared and discussed. Experiments on a mobile robot used in the RoboCup@Work league are discussed as well. The system showing the best performance is capable of estimating the motion with drift as small as 1 under special conditions, though it has been proven to be robust against shakey motion and moderately non-static scenes. © 2014 Springer-Verlag Berlin Heidelberg.