A Gaussian-Generalized-Inverse-Gaussian Joint-Distribution-Based Adaptive MSCKF for Visual-Inertial Odometry Navigation

The visual-inertial odometry (VIO) navigation system plays an important role in providing accurate localization information in absolute navigation information-denied environments, such as indoors and obstruction-filled scenes. However, the working environment may be dynamic, such as due to illumination variations and texture changing in which case the measurement noise of the camera will be nonstationary, and thereby, the VIO exhibits poor navigation using the fixed measurement noise covariance matrix (MNCM). This article proposes an adaptive filter framework based on the multistate constraint Kalman filter (MSCKF). First, the MNCM is regarded as an identity matrix multiplied by a scalar MNCM coefficient that together with the state vector is jointly modeled as Gaussian-generalized-inverse-Gaussian distributed to achieve adaptive adjustment of the MNCM, from which the proposed adaptive filter framework for the VIO navigation system is derived. The proposed adaptive filter framework can theoretically employ a more accurate MNCM during the filtering and, thus, is expected to outperform the traditional MSCKF. Second, the convergence, computational complexity, and initial parameters influence analyses are given to illustrate the validity of the proposed framework. Finally, simulation and experimental studies are carried out to verify the theoretical and practical effectiveness and superiority of the proposed adaptive VIO filter framework, where the EuRoC datasets testing shows the proposed method is 22% and 29% better than the traditional MSCKF in position and orientation estimation, respectively.