A Robust and Efficient Visual-Inertial Initialization with Probabilistic Normal Epipolar Constraint

TL;DR

This work introduces a robust VIO initialization method that decouples rotation and translation while incorporating uncertainty-aware cues. A Probabilistic Normal Epipolar Constraint (PNEC) is used to refine gyroscope-bias estimation by accounting for 2D feature uncertainty, and a fused IMU-vision pipeline estimates velocity, gravity, and scale, followed by a refinement that improves gravity and scale accuracy. Across EuRoC and TUM VI datasets, the method yields lower gyroscope bias, rotation, gravity, and scale errors than strong baselines, while remaining suitable for real-time operation. The approach enhances initialization robustness in challenging motions, enabling more accurate and reliable VIO in practical scenarios.

Abstract

Accurate and robust initialization is essential for Visual-Inertial Odometry (VIO), as poor initialization can severely degrade pose accuracy. During initialization, it is crucial to estimate parameters such as accelerometer bias, gyroscope bias, initial velocity, gravity, etc. Most existing VIO initialization methods adopt Structure from Motion (SfM) to solve for gyroscope bias. However, SfM is not stable and efficient enough in fast-motion or degenerate scenes. To overcome these limitations, we extended the rotation-translation-decoupled framework by adding new uncertainty parameters and optimization modules. First, we adopt a gyroscope bias estimator that incorporates probabilistic normal epipolar constraints. Second, we fuse IMU and visual measurements to solve for velocity, gravity, and scale efficiently. Finally, we design an additional refinement module that effectively reduces gravity and scale errors. Extensive EuRoC dataset tests show that our method reduces gyroscope bias and rotation errors by 16\% and 4\% on average, and gravity error by 29\% on average. On the TUM dataset, our method reduces the gravity error and scale error by 14.2\% and 5.7\% on average respectively. The source code is available at https://github.com/MUCS714/DRT-PNEC.git

Figures (3)

• Figure 1: Geometry of the normal epipolar constraint (NEC) and the relationship between gyroscope bias and NEC. The normal vectors $\mathbf{n}_{1}$ and $\mathbf{n}_{2}$ are perpendicular to the epipolar plane where $\mathbf{f}_{i}^{1}$($\mathbf{f}_{i}^{2}$) and $\mathbf{f}_{j}^{1}$($\mathbf{f}_{j}^{2}$) are located (red and green), and all normal vectors are in the same plane (yellow), forming a constraint that can be used to solve the rotation $\mathbf{R}_{c_{i} c_{j}}$ (orange). The problem of solving $\mathbf{R}_{c_{i} c_{j}}$ is transformed into the problem of solving the gyroscope bias (pink) by using the extrinsic parameter $\mathbf{R}_{bc}$ (blue).
• Figure 2: Gyroscope bias errors and rotation RMSE on EuRoC sequences.
• Figure 3: Angular velocity and scale error visualizations for the MH02 dataset and MH03 dataset. The first-row image is MH02, and the second-row image is MH03. The column (a) shows the trajectory of the corresponding dataset colored based on the angular velocity. The columns (b) and (c) show the trajectory of our method and the DRT-t method based on the scale error colored on the corresponding dataset, respectively. The scale error is between 0 and 1. The lighter the color, the smaller the error.