Comparison of Visual-Inertial Sensor Fusion Techniques for GNSS-Resilient Navigation

This paper analyzes and compares different visual-inertial (V-INS) sensor fusion strategies, conceived to enable GNSS-resilient navigation of Unmanned Aerial Vehicles (UAV). The strategies share the same data fusion architecture, based on a loosely coupled Extended Kalman Filter (EKF), but have variable characteristics concerning the image processing and pose estimation pipelines. Specifically, four different geolocation strategies are investigated to retrieve range information about the detected visual features in the observed scene, exploiting a flat Earth assumption, a local flat Earth assumption, a Digital Elevation Model (DEM) intersection and a triangulation method, respectively. The latter approach, in particular, does not require any geometrical assumption or any a priori knowledge on the observed scene. Furthermore, an original approach to provide the EKF with a consistent estimate of the covariance associated to the visual-based pose solution is proposed. The navigation performance is evaluated by means of numerical simulations reproducing a straight trajectory with hovering and a curved trajectory, both over an almost flat and a mountain scenario. In the best case of the flat scenario, the percentage error on the position estimate, computed with respect to the UAV traveled distance, is less than 0.15%, and the error on the attitude angles estimate is less than 0.08°, while the Root Mean Square (RMS) values of the same quantities are about 0.09% and 0.04°, respectively. In the mountain scenario the performance is degraded but the maximum errors are still kept approximately within 1% and 1o, respectively.