Real-time Visual-inertial Localization using Summary Maps

Localization in a global reference frame constitutes a fundamental milestone towards high-level applications in robotics such as autonomous navigation and obstacle avoidance. Visual-inertial SLAM became a compelling approach for this task, despite its inherent drift and local pose estimation. A solution to these shortcomings, however, can be achieved by matching against maps built in previous sessions. Commonly, a careful data selection is performed to keep the map size traceable and thus enable localization in real-time. Although, such a map summarization usually guarantees global localization coverage, the accuracy suffers due to fewer matches. In this work, we aim at mitigating this effect by directly integrating the scarce 2d-3d matches with visual feature tracks and inertial measurements in the framework of a slidingwindow based optimization. We compare our approach to motion tracking data and demonstrate that such a joint estimation yields smoother and more accurate global pose estimates than related methods that loosely integrate 6-DoF localization poses with VIO. Finally, we evaluate the impact of varying map summarization parameters on the trade-off between map-size and localization accuracy and demonstrate that our approach allows for a more aggressive summarization while retaining the robustness and accuracy achieved with larger maps.