Vision-Based Topological Localization for MAVs

摘要

Vision-based topological localization is recently emerging as a promising alternative to metric pose estimation techniques in robotic navigation systems. Contrarily to the latter, which suffer from a quick degradation of their performance under non-ideal conditions (e.g., scenes with poor illumination and low amount of textures), topological localization trades off precise metric positioning with a more robust and higher-level location representation. State-of-the-art works in this direction, however, often neglect the spatiotemporal relationships between poses that are naturally induced by robotic navigation. Furthermore, these techniques are nearly unexplored for autonomous flying platforms. Inspired by these considerations, in this work, we propose a vision-based topological localization approach designed for Micro Aerial Vehicles (MAVs) applications. Our strategy exploits the framework of graph recurrent neural networks to model the spatial and temporal dependencies and estimate the location of the robot with respect to a topological graph representing the environment. We demonstrate with experiments on different sets of scenarios, including scenes that considerably differ from those used in the training phase, that our approach is able to outperform state-of-the-art place recognition baselines.