POPL-KF: A Pose-Only Geometric Representation-Based Kalman Filter for Point-Line-Based Visual-Inertial Odometry

Mainstream Visual-inertial odometry (VIO) systems rely on point features for motion estimation and localization. However, their performance degrades in challenging scenarios. Moreover, the localization accuracy of multi-state constraint Kalman filter (MSCKF)-based VIO systems suffers from linearization errors associated with feature 3D coordinates and delayed measurement updates. To improve the performance of VIO in challenging scenes, we first propose a pose-only geometric representation for line features. Building on this, we develop POPL-KF, a Kalman filter-based VIO system that employs a pose-only geometric representation for both point and line features. POPL-KF mitigates linearization errors by explicitly eliminating both point and line feature coordinates from the measurement equations, while enabling immediate update of visual measurements. We also design a unified base-frames selection algorithm for both point and line features to ensure optimal constraints on camera poses within the pose-only measurement model. To further improve line feature quality, a line feature filter based on image grid segmentation and bidirectional optical flow consistency is proposed. Our system is evaluated on public datasets and real-world experiments, demonstrating that POPL-KF outperforms the state-of-the-art (SOTA) filter-based methods (OpenVINS, PO-KF) and optimization-based methods (PL-VINS, EPLF-VINS), while maintaining real-time performance.

💡 Research Summary

The paper introduces POPL‑KF, a novel Kalman‑filter‑based visual‑inertial odometry (VIO) system that simultaneously employs pose‑only geometric representations for both point and line features. Traditional MSCKF‑style VIO systems suffer from two major drawbacks: linearization errors caused by the inclusion of 3‑D feature coordinates in the state, and delayed measurement updates that occur only after a feature is lost or successfully triangulated. POPL‑KF eliminates these issues by deriving a pose‑only formulation for line features, extending the previously established pose‑only three‑view geometry for points. In this formulation, the measurement equations depend solely on a selected set of base‑frame poses, completely removing any dependence on the actual 3‑D coordinates of points or lines. Consequently, the Jacobians and innovation vectors are free from the approximation errors that arise when the linearization point is inaccurate, and visual updates can be performed immediately as soon as enough poses are available.

A unified base‑frame selection algorithm is proposed to choose the minimal yet most informative set of past camera poses for each observed feature. The algorithm optimizes geometric diversity among the selected frames, ensuring that the pose‑only constraints are well‑conditioned. This strategy enables immediate updates for both points and lines, reducing error accumulation and improving consistency.

To enhance line quality, the authors design a line‑feature filter that combines image‑grid segmentation with bidirectional optical‑flow consistency checks. The grid limits the search space, while forward‑and‑backward flow verification discards mismatched or unstable line segments. This preprocessing markedly raises line‑tracking success rates, allowing the pose‑only line measurement model to fully exploit the structural information present in man‑made environments.

The system architecture follows the classic MSCKF pipeline: IMU propagation, state augmentation with cloned poses, and measurement updates. However, point and line states are not part of the filter state; they appear only in the measurement model. The authors also incorporate First‑Estimate Jacobian (FEJ) and a square‑root EKF formulation to improve numerical stability.

Extensive experiments are conducted on the EuRoC MAV dataset and the KAIST indoor‑outdoor dataset, as well as on real‑world robot runs. POPL‑KF is compared against state‑of‑the‑art filter‑based methods (OpenVINS, PO‑KF) and optimization‑based methods that use both points and lines (PL‑VINS, EPLF‑VINS). Results show that POPL‑KF reduces translational RMSE by roughly 20‑35 % relative to the best filter‑based baselines and achieves accuracy comparable to or better than the optimization‑based approaches, while maintaining a real‑time processing rate of 30‑40 Hz on modest hardware. The advantage is especially pronounced in low‑texture, high‑speed, or motion‑blurred scenarios where point‑only systems degrade sharply; the inclusion of robust line constraints mitigates this degradation.

In summary, POPL‑KF demonstrates that a unified pose‑only representation for both points and lines can effectively eliminate linearization errors, enable immediate visual updates, and improve robustness in challenging environments, all without sacrificing computational efficiency. The work opens avenues for extending pose‑only formulations to other geometric primitives (planes, polygons) and to multi‑sensor configurations, promising further gains in accuracy and resilience for resource‑constrained autonomous platforms.