Outlier-Robust Extended Kalman Filtering for Bioinspired Integrated Navigation System

This article proposes a set of novel attitude estimation algorithms for a bioinspired integrated navigation system including an insect-inspired polarization sensor, an inertial measurement unit, and a global navigation satellite system (GNSS). In particular, the biomimetic polarization sensor introduced as the new type of navigation component has an orientation performance comparable to that of a magnetometer. Moreover, this small and light sensor is immune to magnetic interference, making it highly suitable for orientation in unmanned aerial vehicles (UAVs). However, inaccurate modeling, unknown parameters, and unknown statistical characteristics of noise are common problems with the orientation solution of the sensor. Therefore, we theoretically compare various adaptive and robust filtering algorithms, select the best error prediction covariance and noise covariance correction schemes, introduce Kalman smoothers and approximate inference methods, and then propose two outlier-robust algorithms. This approach can suppress multiple outliers from the two steps of prediction and estimation. The polarization sensor and GNSS update the state and error covariance according to the asynchronous fusion mode. The prediction error covariance, the polarization measurement noise covariance, and the GNSS measurement noise covariance are approximately estimated by the variational inference method. To validate the performance of our algorithms, we collected a set of multi-rotor UAV flight data for testing. Experimental results show that the proposed algorithms have better robustness performance than existing state-of-the-art algorithms. Note to Practitioners —This article was motivated by Bioinspired autonomous navigation systems suffering from inaccurate modeling, unknown parameters, and noise with unknown statistical properties. Subject to the large uncertainty in the flight of the UAV, the traditional filtering algorithm has slow convergence or even divergence when processing such actual data. In this article, we first establish a unified state estimation framework. Then, we mathematically describe correction strategies for prediction error covariance and measurement noise covariance. Considering the need for real-time estimation, we adopted the extended Kalman filter. However, this introduces a truncation error, then, Rauch-Tung-Striebel smoothing is introduced to improve the linearized reference point. However, the role of robust filtering is limited, and the non-Gaussian state still exists. Finally, the approximate inference method is used to estimate the state. The proposed method integrates adaptability, robustness, reasoning, and smoothing, and is suitable for dealing with multi-source interference problems in multi-source information fusion.