Unsupervised matching of visual landmarks for robotic homing using Fourier-Mellin transform

This paper presents an algorithm that uses visual information to achieve the homing of an autonomous agent inside a previously visited environment. An image grabbed at the target position is compared with the currently perceived one to determine the relative position of the robot and of its target. Only particular regions of the image, called Visual References, are taken into account. A Visual References correlation criterion that uses the Fourier–Mellin Transform to match the Visual References in different images is employed. This transform in fact allows computing Visual References that are invariant to rotation, scaling and translation. Robustness due to the use of the Mellin Transform in the Visual References selection and coupling leads to more precise navigation. Tests and results are presented.