Improving Attitude Estimation Using Gaussian-Process-Regression-Based Magnetic Field Maps

Abstract

Magnetometers measure the local magnetic field and are present in most modern inertial measurement units (IMUs). Readings from magnetometers are used to identify Earth’s Magnetic North outdoors, but are often ignored during indoor experiments since the magnetic field does not behave how most expect. This paper presents methods to create, validate, and visualize three-dimensional magnetic field maps to expand the use of magnetic fields as a sensing modality for navigation. The utility of these maps is measured in their ability to accurately represent the magnetic field and to enable dynamic attitude estimation. In experiments with motion capture truth data, a small multicopter with three-axis inertial measurements, including magnetometer, traversed five flight profiles distinctly exciting roll, pitch, and yaw motion to provide interesting trajectories for attitude estimation. Indoor experimental results were compared to those outdoors to emphasize how spatial variation in the magnetic field drives the need for our mapping techniques. Our work presents a new way of visualizing 3D magnetic fields, which allows users to better reason about the magnetic field in their workspace. Next, we show that magnetic field maps generated from coverage patterns are generally more accurate, but training such maps using observations from desired flight paths is sufficient in the vicinity of these paths. All training sets were interpolated using Gaussian process regression (GPR), which yielded maps with <1 μT of error when interpolating between and extrapolating outside of observed locations. Finally, we validated the utility of our GPR-based maps in enabling attitude estimates in regions of high magnetic field spatial variation with experimental data.