TL;DR
This paper introduces a novel multi-layer framework for robust, long-term global positioning by fusing visual, inertial, and GNSS data. It addresses the challenge of drift in visual-inertial systems, especially in complex environments, by introducing a hierarchical approach and a new MAP-based initialization method that significantly improves accuracy and speed.
What Problem Does This Paper Address?
Accurate and reliable 6-DOF (six degrees of freedom) camera pose estimation is crucial for many applications like robotics, autonomous driving, and augmented reality. While Visual SLAM (Simultaneous Localization and Mapping) is cost-effective, it suffers from scale uncertainty and drift over time. Visual-Inertial (VI) systems improve this by using IMU data to recover scale and correct IMU bias, but they still accumulate errors in the long term, limiting their use in large-scale scenarios. Fusing GNSS (Global Navigation Satellite System) data offers global positioning information, but integrating it effectively with VI systems, especially during GNSS signal degradation or initialization, remains a challenge. Existing VI-GNSS fusion methods often struggle with accurate initialization, robustness to sensor noise, and efficient handling of different data frequencies.
Methods
The paper proposes a multi-layer VI-GNSS global positioning framework designed for rapid initialization and robust online localization.
- Inner-Layer: Tightly-Coupled Visual-Inertial-GNSS BA: This layer focuses on local motion estimation. It uses a sliding window Bundle Adjustment (BA) to tightly fuse visual, inertial (IMU), and GNSS velocity measurements. The key innovation here is using GNSS velocity to accurately estimate metric scale and correct IMU bias, mitigating scale drift inherent in visual odometry. This layer also calibrates extrinsic parameters between the camera and IMU.
- Outer-Layer: Loosely-Coupled Pose Graph Optimization: This layer handles global localization over longer periods. It fuses the local estimates from the inner layer with GNSS position and course data. This loosely-coupled approach is designed to be robust to GNSS degradation and outliers, and it performs a 4-DOF pose graph optimization (focusing on yaw for global alignment) to correct accumulated drift and ensure alignment with the world frame.
- MAP-based Initialization: A novel Maximum a Posteriori (MAP)-based initialization framework is introduced. It first estimates initial poses with metric scale using a sliding window VO aided by GNSS velocity. Then, it numerically solves for the world frame extrinsic parameters. Finally, it refines all state variables (IMU bias, scale, extrinsic parameters, etc.) using a MAP optimization over the first few frames. This approach aims for faster, more accurate initialization compared to purely analytical or optimization-based methods, avoiding local optima.
Datasets
The proposed framework is evaluated on two public datasets:
- EuRoC MAV Dataset [28]: Used for evaluating the performance of the entire framework, particularly the initialization, on 11 indoor sequences.
- Kaist Urban and Highway Dataset [29]: Used to validate the long-term practicability in challenging outdoor scenarios, including GNSS degradation, sparse features, and low IMU excitation.
The authors also conducted tests on a self-developed intelligent vehicle platform.
Key Insights and Ideas
- GNSS Velocity for Scale Correction: Using GNSS velocity in the inner-layer BA is a critical insight for accurately recovering metric scale and correcting drift in visual odometry, especially when GNSS position data might be noisy or unavailable.
- Hierarchical Fusion for Robustness: The multi-layer approach effectively decouples local motion estimation from global pose alignment. The inner layer handles short-term accuracy and scale, while the outer layer manages long-term drift and global consistency, leading to enhanced robustness.
- MAP Initialization for Speed and Accuracy: The proposed MAP-based initialization, combining analytical solutions with optimization, provides a strong initial guess for all state variables, leading to faster convergence and more accurate results than existing methods, particularly in avoiding local minima.
- 4-DOF Pose Graph for Global Alignment: Focusing the outer-layer optimization on 4-DOF (primarily yaw) leverages the strengths of GNSS for global alignment while mitigating potential issues with roll and pitch estimation from GNSS alone.
Key Results
- Initialization Performance: The proposed method achieved significantly better scale errors (average 8.13%) compared to VINS-Mono (average 25.90%) on the EuRoC dataset.
- Localization Accuracy:
- On EuRoC, the proposed method achieved an average translation error of up to 63% lower than VINS-Fusion and other state-of-the-art methods.
- On the Kaist dataset, the localization error was reduced by 63% compared to VINS-Fusion on their self-developed intelligent vehicle platform.
- Robustness: The framework demonstrated improved robustness to GNSS degradation and challenging outdoor scenarios (urban canyons, sparse features, low IMU excitation).
Limitations and Future Work
- Initialization Time: While improved, the initialization process still requires a longer time (tTot) than purely visual-inertial initialization methods because it includes the crucial step of local-global frame alignment.
- GNSS Frequency: The paper notes that GNSS data has a low frequency (1Hz) and is not always synchronized, requiring interpolation. This can still be a limiting factor.
- Future Work: The authors suggest applying the algorithm to other VI-GNSS and odometer systems, implying its generalizability. Further improvements could focus on even faster initialization or more seamless integration of low-frequency GNSS data.
Summary updated