Floor Texture Visual Servo Using Multiple Cameras for Mobile Robot Localization

Abstract

A local localization algorithm for mobile robots has been proposed, which is based on the idea of using multiple off-the-shelf webcams to perform ground texture tracking. The localization module has been developed on a custom built robot and tested in real indoor environments with dramatic improvement over encoder based dead reckoning approaches. To take advantage of the constraints provided by the system and the type of environment the robot is exposed to, various characteristics of the camera were configured and adjusted to reduce the complexity in the tracking task. There are two constraints that are used for the proposed approach to work, which are: The elevation of the camera to the ground remains constant, and The features being tracked can only translate and not rotate in between frames. Due to the processing requirement, only two filters are actively used, which are the lens warp removal filter and block removal filter. After exploring several scoring algorithms to find the feature, a simple algorithm based on the standard deviation has been used with a shape of 16 by 16 pixel square. To improve the processing time for finding the feature, a prediction is made to where the feature is located, followed by a spiral search sequence to quickly find the best candidate, which has lead to approximately 30% speed up. By accounting for some of the sub-pixel motions by interpolating around the best candidate, the precision of the tracking increased by approximately 6 times. To distinguish between translation and rotation of the robot, a second tracker was introduced to form a two-cameras-on-the-side configuration. The two motion vectors were smoothed by using a sliding window of size 4 and a quadratic weight decay function to better synchronise the two data sources. A hybrid motion model has been introduced to handle two types of motions; regular motion based on the locomotive constraints and irregular motion, caused by bumps and sudden slippages. By switching between the two, the performance of the algorithm showed some improvements even though the frequency of erroneous tracking is already quite small. The proposed localization algorithm has been tested on various surfaces types that are commonly found in indoor environments with less than 1% error on both translation and rotation. It was found that the algorithm did not operate so well on very dark surfaces with highly repetitive or indistinguishable texture patterns. As long as the constraints can be maintained, the approach allows for an immediate and precise localization with low cost hardware at a reasonably small processing cost.