PO-KF: A Pose-Only Representation-Based Kalman Filter for Visual Inertial Odometry
Liqiang Wang, Hailiang Tang, Tisheng Zhang, Yan Wang, Quan Zhang, Xiaoji Niu
Abstract
Visual-inertial state estimation is widely employed in the Internet of Things, with filter-based visual-inertial odometry (VIO) being a popular algorithm due to its balance between computational efficiency and localization accuracy. However, the localization performance of the commonly used multi-state constraint Kalman filter (MSCKF)-based VIO is suffering from linearization errors in feature three-dimensional (3D) positions and delayed measurement updates. Targeting more accurate and robust localization, we incorporate the pose-only representation into the filter-based VIO and propose a pose-only representation-based Kalman filter (PO-KF) in this paper. Leveraging the decoupling of camera poses and feature positions in the pose-only representation, the proposed PO-KF explicitly eliminates feature 3D coordinates from its measurement equation. As a result, the linearization errors caused by feature positions can be removed efficiently, while immediate updates of visual measurements can be conducted. We also introduce an information matrix-derived base-frame selection algorithm to identify the most suitable base-frames for each feature. Extensive experiments on multiple datasets demonstrate that PO-KF outperforms state-of-the-art VIO systems. Notably, PO-KF achieves nearly a 50% reduction in relative pose errors compared to MSCKF-based VIO. Further experiments demonstrate that PO-KF also exhibits superior robustness while maintaining real-time performance comparable to MSCKF-based VIO.