In this section, we evaluate the KVO localization approach and compare it with the A.R.T DTrack system as ground truth. For the evaluation, different movement patterns are performed on the Robotino (see Figure 1) which is localized using odometry, vision, KVO and DTrack.
For the vision localization, the markers are placed on a grid with 70cm edge length and all markers have the same orientation. This results in 2-3 markers in the camera frame for most of the time. For the localization task, we have hard environment conditions. The lamps on the ceiling are turned on, which increases the difficulty of correct marker detection. Our robot has been used for several years, hence mechanical wear and tear becomes noticeable. The robot also has a high, unsymmetrical setup which displaces the center of mass unfavorably. This leads to inaccurate odometry information. In some cases, vision pose estimates are faulty because a marker has been confused with another marker, or a structure on the ceiling has been interpreted as visual marker by mistake. Such faulty vision pose estimation are detected using the state uncertainty of the Kalman-Filter. If the position or orientation of the vision estimation is outside of the 2σ region, as described in Section III-B, the estimation is assumed to be corrupted and ignored. For the Kalman-Filter, the noise matrices Rt and Qt for pose estimation are set beforehand. We assume x and y errors to be independent with standard deviations of 10cm in each direction for vision estimation, and a standard deviation of 5cm for odometry estimation. The standard deviation of the angle is set to 1.5◦ for vision and 0.75◦ for odometry.