ABSTRACT
Localization is fundamental to enable the use of autonomous mobile robots. In this work, we use magnetic-based localization. As Earth's geomagnetic field is stable in time and is not affected by nonmagnetic materials, such as a large number of people in the robot's surroundings, magnetic-based localization is ideal for service robotics in supermarkets, hotels, etc. A common approach for magnetic-based localization is to first create a magnetic map of the environment where the robot will be deployed. For this, magnetic samples acquired a priori are used. To generate this map, the collected data is interpolated by training a Gaussian Process Regression model. Gaussian processes are nonparametric, data-drive models, where the most important design choice is the selection of an adequate kernel function. These models are flexible and generate mean predictions as well as the confidence of those predictions, making them ideal for their use in probabilistic approaches. However, their computational and memory cost scales poorly when large datasets are used for training, making their use in large-scale environments challenging. The purpose of this study is to: (i) enable magnetic-based localization on large-scale environments by using a sparse representation of Gaussian processes, (ii) test the effect of several kernel functions on robot localization, and (iii) evaluate the accuracy of the approach experimentally on different large-scale environments.