GPGM-SLAM: Towards a Robust SLAM System for Unstructured Planetary Environments with Gaussian Process Gradient Maps

Abstract

Simultaneous Localization and Mapping (SLAM) in unstructured planetary environments is a challenging task for mobile robots due to the appearance and structure of the environment. In urban and man-made scenarios, individual objects (e.g. cars, trees or buildings) are easily discernible and the visual appearance is likely to provide unique cues for the purpose of localization. Contrarily, planetary scenarios are often characterized by repetitive structures and ambiguous terrain features. To provide robust place recognition abilities in the context of submap-based stereo visual SLAM, we propose to utilize the gradient of elevation maps generated by Gaussian Processes (GPs). Visual features computed on GP Gradient Maps (GPGMaps) provide means for efficient place recognition, through encoding in Bag-of-Words vectors, and for SE(2) alignment to establish loop closure constraints in a pose graph. We evaluate the proposed SLAM system on relevant Moon-like environments through real data captured on Mt. Etna, Sicily.