DOGE: An Extrinsic Orientation and Gyroscope Bias Estimation for Visual-Inertial Odometry Initialization

TL;DR

The paper tackles the fragility of visual‑inertial odometry initialization when rotational extrinsics drift due to deformation. It introduces DOGE, a rotation‑focused framework that jointly estimates extrinsic orientation and gyroscope bias within normal epipolar constraints, aided by two weighting strategies and a chi‑square failure test, followed by a MAP refinement within an IESKF before translation parallax occurs. Key contributions include a rotation‑only NEC formulation, iterative reweighting with lambda and Feature‑Pairs strategies, and MAP‑based refinement that yields faster and more reliable initialization, demonstrated on EuRoC with significant robustness gains over state‑of‑the‑art methods and competitive efficiency. The approach enables more robust AR/VR and robotic VIO bootstrapping in real‑world conditions where hardware deformation affects extrinsics, particularly rotation, and shows favorable impact when integrated into full VIO pipelines.

Abstract

Most existing visual-inertial odometry (VIO) initialization methods rely on accurate pre-calibrated extrinsic parameters. However, during long-term use, irreversible structural deformation caused by temperature changes, mechanical squeezing, etc. will cause changes in extrinsic parameters, especially in the rotational part. Existing initialization methods that simultaneously estimate extrinsic parameters suffer from poor robustness, low precision, and long initialization latency due to the need for sufficient translational motion. To address these problems, we propose a novel VIO initialization method, which jointly considers extrinsic orientation and gyroscope bias within the normal epipolar constraints, achieving higher precision and better robustness without delayed rotational calibration. First, a rotation-only constraint is designed for extrinsic orientation and gyroscope bias estimation, which tightly couples gyroscope measurements and visual observations and can be solved in pure-rotation cases. Second, we propose a weighting strategy together with a failure detection strategy to enhance the precision and robustness of the estimator. Finally, we leverage Maximum A Posteriori to refine the results before enough translation parallax comes. Extensive experiments have demonstrated that our method outperforms the state-of-the-art methods in both accuracy and robustness while maintaining competitive efficiency.

Figures (6)

• Figure 1: Illustration of the advantage of the proposed method for AR applications: The top row of images shows the simulation scene and the motion trajectory of the camera. In the first 25 seconds, the camera has only pure rotational motion. The detailed configuration can be found in Sec. \ref{['sec:vio_test']}. The bottom left image shows the estimation errors of rotational extrinsic parameters and gyroscope bias. It can be seen that our method leverages the information from the pure rotation stage and gets a quicker convergence. The bottom right image shows the impact of initialization performance on VIO in AR applications. The magnitude of the difference between the virtual object "DOGE" rendered with the estimated pose (purple) and the ground truth (green) in the figure indicates the performance of each method, clearly demonstrating the effectiveness of our method in terms of robustness, accuracy, and low latency.
• Figure 2: Accuracy evaluation: The proposed method significantly outperforms the classical and state-of-the-art methods when the extrinsic orientation is poor.
• Figure 3: Robustness evaluation: The top bar is colored according to the magnitude of the average angular velocity of the data segments (darker colors indicate higher magnitudes). The cyan points indicate data segments where a good estimation was obtained ($\bold{b}^I_g$ error is less than 50% and the $\bold{R}_{CI}$ error is less than 5 degrees). The orange points indicate data segments where a bad estimation is detected. The red points indicate segments where the method does not return a failure flag, but the $\bold{b}^I_g$ error exceeds 50% or the $\bold{R}_{CI}$ error is larger than 5 degrees.
• Figure 4: Ablation expriment
• Figure 5: Test on EuRoC: Each methods adopt the same configuration as VINS-Mono. Due to ignoring gyroscope bias, VINS-Mono(Pre) performs worse than VINS-Mono(Post). On contrary, the proposed method gets an outstanding performance on almost all sequences.