FeRF-BEVIO: A Feature Radiance Field-Based Bidirectionally Enhanced Visual-Inertial Odometry
Yuanxi Gao, Jing Yuan, Shizhuo Yu, Xuebo Zhang · IEEE Transactions on Automation Science and Engineering · 2025
To address the issues of inaccurate feature association and insufficient utilization of the complementarity between the visual and inertial information in visual-inertial navigation systems, this paper proposes a feature radiance field (FeRF)-based bidirectionally enhanced visual-inertial odometry (FeRF-BEVIO). FeRF is a novel method for describing environment features, incorporating both the grayscale and position of point and line features. It provides a unified framework for representing and storing features, regardless of whether their depth has converged. Then, a bidirectional visual-inertial enhancement method is designed based on FeRFs. Specifically, on one hand, the information of the inertial measurement unit (IMU) is utilized to aid visual feature association within FeRFs. On the other hand, the visual information is employed to refine the IMU parameter estimation. This bidirectional enhancement process is iterated to improve the integration of the visual and inertial data. At last, the system jointly optimizes the robot poses and point-line feature parameters within a sliding window and updates FeRFs accordingly. Comparative experiments on public datasets and in the real-world environments demonstrate that FeRF-BEVIO outperforms state-of-the-art visual-inertial odometry (VIO) systems in both accuracy and robustness. Therefore, FeRF-BEVIO is highly suitable for navigation and simultaneous localization and mapping (SLAM) of micro aerial vehicles (MAVs).