CNN-Derived Local Features for Condition-Invariant Robot Localization with a Single RGB Sensor

A conditio sine qua non for any autonomous system refers to its ability to localize itself into a known environment. Towards this end, camera sensors are typically deployed for measuring the relative transformation between the mobile platform and a pre-mapped scene due to their low cost and substantial accuracy. This paper deals with the task of robot localization under different environmental conditions using a single RGB sensor. To achieve this, we diverge from conventional approaches that detect local points of interest using hand-crafted sets of rules, and we utilize the cognitive properties of modern deep learning models for feature detection and description. The proposed architecture is evaluated and compared against other traditional techniques under a wide range of environmental conditions that significantly alter the view of a previously recorded area.

Abbreviations: Autonomous Systems; Robot Localization; Deep Learning; Local Features