Matching geometry for long-term monocular camera localization

Localizing a camera with respect to a given map is a fundamental problem in vision-based navigation. Realworld applications require methods that are capable of longterm operation because recordings of available maps often date back considerably compared to the time of localization. Unfortunately, the photometric appearance of the environment can change tremendously even over short periods, making image matching a difficult problem. Since geometric properties of the environment are typically more stable than its photometric characteristics, we propose to rely on matching geometry in order to achieve long-term camera localization. Thus, our approach is agnostic to changes in photometric appearance. We present real-world experiments which demonstrate that our method accurately tracks the 6-DoF pose of a camera over long trajectories and under varying conditions. We evaluate our method using publicly available and own datasets and visualize the successful tracking by post-colorizing the given geometric map.