Experimental Comparison of Odometry Approaches

Odometry is an important input to robot navigation systems, and we are interested in the performance of vision-only techniques. In this paper we experimentally evaluate and compare the performance of wheel odometry, monocular feature-based visual odometry, monocular patch-based visual odometry, and a technique that fuses wheel odometry and visual odometry, on a mobile robot operating in a typical indoor environment. 1 Motivation and Problem Statement Fig. 1. The Adept Guiabot robot used in this paper. Today most research robots, the prototypes of future commercial systems, use laser rangefinders (LRFs) as the sensor that informs the essential parts of a navigation system: odometry, place recognition and mapping. LRFs are self-contained, reliable and provide metric information about the world, but since they are based on mature electro/optical/mechanical technologies will always be more expensive than cameras. They also only capture just a slice through the scene whereas a camera can instantaneously capture a view of a whole area which is important for fast moving robots. Imaging 3D sensors such as flash LIDAR (eg. Swiss Ranger) or Kinect technology are entirely solid-state but function poorly in the presence of daylight. Theoretically vision is able to perform the tasks of odometry and place recognition, and the sensors are in principle cheaper and smaller than LRFs. Vision has other advantages: it is a passive sensor and thus immune from the interference problems that will occur when multiple robots emitting infra-red light are operating in close proximity; it provides rich scene information such as color and texture which can assist semantic tagging; and when combined with motion it can generate 3D scene structure. However vision is rarely reported for indoor navigation and this paper details early results from a larger project aimed at developing a purely vision-based large-scale and long-term mobile robot navigation system that can work both indoors and outdoors. This paper is concerned specifically with maximizing the accuracy of odometry; we view odometry as an essential part of any navigation system and recognize that external mechanisms to “close-the-loop” can improve odometry performance significantly, but the focus here is on maximizing odometry performance in isolation. We compare the performance of several odometry approaches: raw wheel odometry, patch based monocular visual odometry (VO), state-of-the-art feature-based monocular VO, a method of fusing wheel odometry and VO using confidence metrics generated from the VO, and laser scan matching which we refer to as ground truth. The performance of these odometry techniques is evaluated in two distinct indoor environments that present a number of different challenges.