An accurate, up-to-date estimation of it's own pose is important for any mobile robot that should be able to navigate in and/or interact with the physical world.
While there are many methods of pose estimation, they are often lacking in one or more specific aspects.
In particular, they may be inaccurate, low-rate, provide incomplete data or require external architecture.
In my Master's thesis, I tackle the problem of pose estimation for the robotic rover Artemis by integrating a sensor fusion component, seeking to combine the advantages of multiple pose estimators for a single, accurate, high-rate pose estimation.
Further, a visual odometry component is added as an input to said estimator, allowing the use of cameras for pose estimation. I document the steps used in the evaluation of multiple existing visual odometry/SLAM components for this use-case.
Additionally, I present the results of said evaluation and justify my final selection.
Further, I give an overview over the Inavriant Extended Kalman Filter that is used as pose fusion component and present early results on a self-recorded dataset.
I also detail how I expect to modify said filter to estimate the inherent scale of the poses estimated by the (monocular) visual odometry component.
Finally, I outline further developments in the scope of the thesis. This includes the integration of additional pose sources like landmarks and wheel odometry.
It also includes the integration of the filter into Artemis's software stack, while allowing possible use on other robots. Lastly, a final evaluation shall evaluate the efficacy of the resulting pose-estimation stack in different scenarios and with different sensor configurations.