A context-aware framework for visual localization

The purpose of this project is to create a visual localization system – a system that identifies where it is in the world – that can operate autonomously for long time periods and in large-scale environments. The system must continuously assess if the appearance of the environment has changed and if many places in the environment look similar. It can then adapt its underlying models of the world accordingly.The project will be organized and undertaken at Örebro University and focus on the completion of three scientific goals. The first milestone (Apr 2020) will define a metric that quantifies how much the environment has changed and how distinctive it is. The second milestone (Dec 2021) will build a localization system that adapts to the level of environmental change. The final milestone (Dec 2022) will improve the localization reliability by integrating geometric information about the world into the localization evaluation.The localization system will use a probabilistic framework and its knowledge about the environment will be encoded as likelihood models that dynamically learn and update as the system moves through the world. It will be tested on a mobile robot operating outdoors over multiple days, during different seasons, and in variable weather conditions.This system will extend the long-term reliability of visual localization systems and provide a fundamental step towards mobile systems that can operate autonomously in real-world environments.