Evaluating Map-Based RGB-D SLAM on an Autonomous Walking Robot

This paper demonstrates an application of a Simultaneous Localization and Mapping algorithm to localize a six-legged robot using data from a compact RGB-D sensor. The algorithm employs a new concept of combining fast Visual Odometry to track the sensor motion, and a map of 3-D point features and robot poses, which is then optimized. The focus of the paper is on evaluating the presented approach on a real walking robot under supervision of a motion registration system that provides ground truth trajectories. We evaluate the accuracy of the estimated robot trajectories applying the well-established methodologies of Relative Pose Error and Absolute Trajectory Error, and investigate the causes of accuracy degradation when the RGB-D camera is carried by a walking robot. Moreover, we demonstrate that the accuracy of robot poses is sufficient for dense environment mapping in 3-D.