Robust Gaussian Filtering

Most widely used state estimation algorithms, such as the Extended Kalman Filter and the Unscented Kalman Filter, belong to the family of Gaussian Filters (GFs). Unfortunately, GFs fail for observation models described by a fat-tailed distribution. This is a serious limitation because thin-tailed observation models, such as the Gaussian distribution, are sensitive to outliers. In this paper, we show that GFs can be enabled to work with fat-tailed models by applying a feature function to the measurement. Furthermore, we find an optimal feature function using variational inference techniques. Simulation results demonstrate the robustness of the proposed filter to outliers in the measurements.