Improving UAV Localization Accuracy by Tracking Visual Features in a Georeferenced Map

AbstractWith the rapid development of computer vision, visual odometry (VO) and visual simultaneous localization and mapping (vSLAM) have been widely used in unmanned aerial vehicles. However, VO suffers from drifts and VSLAM requires loop closure to improve localization accuracy. In this paper, we propose to combine a visual-inertial odometry with a 2D georeferenced map to achieve high precision localization performance in UAVs. Firstly, the georeferenced map is preprocessed to build a visual landmark database. Then a micro inertial measurement unit (MIMU) and a downward facing camera are fused to form a visual inertial odometry (VIO), which estimates relative motion between camera frames. To cure the VIO drifts, we further register the features tracked by the VIO with the 2D georeferenced map. While most conventional methods perform one-shot geo-registration or use 3D maps, we propose to use a 2D geo-referenced map and track the registered features in multiple frames to improve localization accuracy. Finally, a factor graph is built to fuse the relative motion from VIO with the geo-registration information to obtain consecutive and accurate localization results. We valid the proposed algorithm with real flight experiments and the results show significant localization accuracy improvements.KeywordsVIOFactor graph optimizationRegistration with map