MCL-based global localization of cleaning robot using fast rotation-invariant corner matching method

Mobile robot navigation with ceiling features such as a corner which is one of the most popular visual features used in robotics has been widely studied because of its practicality and high performance, and recently low-cost robots have started to use this navigation technique. A cleaning robot is a good example. This study is focused on global localization of a cleaning robot and MCL, one of the popular localization methods, was used with ceiling corners. However, MCL-based global localization is a very time consuming task even on a PC, and so a fast rotation-invariant corner matching method was proposed in this study to reduce the time of global localization with corner features. A pixel-based sum of squared differences (SSD) method has been widely used for corner matching. However, because this method cannot match corners with rotation changes, it is unsuitable for a cleaning robot where corners observed from the robot have rotation changes. In our approach, the image around a corner is divided into some partitions and the representative values of all partitions are computed to generate a rotation-invariant descriptor. This descriptor consists of a small number of values, and two descriptors are simply compared to match two corners. Various experiments on a PC and an embedded system verify that matching by the proposed method is very fast and invariant to a rotation change, and is more suitable for a cleaning robot than the pixel-based SSD method. Moreover, global localization can be conducted using this matching method.