Reliable local odometry is a practical requirement for autonomous agricultural vehicles operating under orchard conditions, where GNSS performance may degrade because of canopy cover, multipath, and other Non-Line-of-Sight effects. A low-cost solution for onboard state estimation is Visual-Inertial Odometry, but it remains susceptible to drift over long trajectories and during turning maneuvers. To address this issue, this paper introduces a lightweight adaptive covariance strategy within the OpenVINS Multi-State Constraint Kalman Filter. The proposed method scales gyroscope process noise according to angular velocity and adjusts visual measurement noise at the feature level using track length and reprojection quality. The resulting estimator is evaluated on three real-world rover runs conducted in an orchard, using PPK-corrected GNSS as reference over the portions where such reference is valid. Across the tested runs, the adaptive variant shows lower local drift than standard OpenVINS, with the clearest gains observed in the longest trajectory and during row-change maneuvers, where translational Relative Pose Error is reduced by approximately 26% relative to standard OpenVINS. Taken together, these results indicate that lightweight online covariance adaptation is an effective practical strategy to improve local visual-inertial estimation under the tested orchard conditions.
Dynamic Drift Compensation for Visual-Inertial Odometry in Unstructured Orchards
Girlanda, Federico
;Osman, Anas;Vecchio, Massimo;Antonelli, Fabio
2026-01-01
Abstract
Reliable local odometry is a practical requirement for autonomous agricultural vehicles operating under orchard conditions, where GNSS performance may degrade because of canopy cover, multipath, and other Non-Line-of-Sight effects. A low-cost solution for onboard state estimation is Visual-Inertial Odometry, but it remains susceptible to drift over long trajectories and during turning maneuvers. To address this issue, this paper introduces a lightweight adaptive covariance strategy within the OpenVINS Multi-State Constraint Kalman Filter. The proposed method scales gyroscope process noise according to angular velocity and adjusts visual measurement noise at the feature level using track length and reprojection quality. The resulting estimator is evaluated on three real-world rover runs conducted in an orchard, using PPK-corrected GNSS as reference over the portions where such reference is valid. Across the tested runs, the adaptive variant shows lower local drift than standard OpenVINS, with the clearest gains observed in the longest trajectory and during row-change maneuvers, where translational Relative Pose Error is reduced by approximately 26% relative to standard OpenVINS. Taken together, these results indicate that lightweight online covariance adaptation is an effective practical strategy to improve local visual-inertial estimation under the tested orchard conditions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
