Low-Rate Cross-View Map Updates for Drift-Bounded Fixed-Wing VIO in GNSS-Denied Missions

Fixed-wing UAVs operating in GNSS-denied conditions require stable state estimates over long distances, yet standalone visual-inertial odometry (VIO) still accumulates drift during sustained forward flight because loop closures are rare and viewpoint variation is limited. This paper presents a drift-bounded navigation architecture with five components: a camera–IMU VIO front end, a motion-aware keyframe selector, a cross-view aerial-map retrieval module, an image-to-map alignment stage, and a confidence-gated conservative correction back end. The VIO front end runs continuously to provide high-rate local pose, while map anchoring is invoked only at low rate on selected keyframes. A map update is accepted only when the retrieval score is high, the geometric inlier ratio is sufficient, the reprojection error is low, and the estimated heading is consistent with the onboard solution; in the present implementation, the corresponding gates are 0.80, 0.35, 3.0 px, and 8°, respectively. Accepted updates correct global position and yaw without resetting the local VIO trajectory.

In representative high-fidelity simulation with staged GNSS loss over a 24 km route, standalone VIO reached 0.92% final drift, whereas the proposed method reduced drift to 0.28%, with an 81% update success rate and 22.4 m absolute trajectory error. In shorter fixed-wing flight tests over 8.6 km, drift decreased from 1.34% to 0.41%, with 74% accepted updates and 34.7 m absolute trajectory error. Compared with continuous map matching, the low-rate policy preserved similar drift behavior while reducing map-matching calls by 75% and sustaining 18 FPS on an Intel NUC 8i7BEH-class onboard computer. The main contribution is a fixed-wing-oriented fusion policy that uses geo-referenced maps as occasional, verified corrections rather than as a continuously active localization source.