Invariant-EKF-Based GNSS/INS/Vision Integration with High Convergence and Accuracy

Nowadays, the tight integration of Global Navigation Satellite System (GNSS), inertial navigation system (INS), and visual odometry has become a prevalent way to obtain continuous and drift-less pose estimation. Unfortunately, convergence, a prerequisite for the estimator to correctly fuse multiple-source information into a single coherent state estimate, remains challenging in complex and various operating environments. Moreover, certain beneficial information, such as GNSS multiple frequency resources, is generally untapped in such systems. To improve accuracy and convergence, we proposed an invariant extended Kalman filter-based framework that tightly couples stereo vision, INS, and GNSS, supporting both precise point positioning and real-time kinematic mode. By maximizing the advantage of triple-band GNSS measurements, the proposed system ensures high-precision pose estimation, shortens initialization process, and raises fixing rate. Meanwhile, the error propagation of the proposed system is log-linear and relatively independent of state prediction, contributing to its observability and convergence, which are examined by both theoretical derivation and experimental Monte Carlo tests. The open-sourced GVINS dataset and a field experiment are used for system evaluation in different GNSS observing environments, and the results indicate that the proposed system outperforms state-of-the-art approaches in both positioning accuracy and convergence capability under large initial perturbations.