Evaluation of a CNN-based Visual Place Recognition system for GPS-denied Navigation of VTOL Vehicles

Abstract

Current Vertical Take-Off and Landing (VTOL) systems rely mainly on Global Positioning System (GPS) for autonomous navigation. Due to the unreliability of GPS, the need for alternative methods has become significant. Among the alternative approaches, Visual Place Recognition (VPR) systems have taken prominence. The latest advancements of these VPR systems involve using deep neural networks, such as Convolutional Neural Nets (CNNs), to overcome the limitations of conventional feature-based systems. These VPR methods have been tested and validated primarily for ground-based datasets. However, to properly assess the suitability of those approaches in VTOL navigation, they need to be evaluated for aerial image data sets. This study evaluates the performance of a CNN-based VPR system against a conventional feature-based method for an aerial image dataset, focusing mainly on the systems' front-end. Furthermore, experimental validation of the CNN-based VPR system is conducted. The results suggest that it is a better addition to the navigation stack of a VTOL vehicle under GPS-denied situations.