[Paper Review] Optimization-Based Visual-Inertial SLAM Tightly Coupled with Raw GNSS Measurements

Jinxu Liu, Wei Gao+1 more|arXiv (Cornell University)|Oct 22, 2020

Robotics and Sensor-Based LocalizationEngineering24 references61 citations

TL;DR

This paper proposes an optimization-based visual-inertial SLAM system that tightly couples raw GNSS measurements—pseudoranges and Doppler shifts—with visual features and IMU preintegration in a sliding window framework. By jointly minimizing reprojection, IMU preintegration, and raw GNSS errors while accounting for sensor asynchrony, the method achieves superior global accuracy in urban canyons and low-rise areas, outperforming state-of-the-art VI-SLAM, GNSS single-point positioning, and loose-coupling approaches on public datasets.

ABSTRACT

Unlike loose coupling approaches and the EKF-based approaches in the literature, we propose an optimization-based visual-inertial SLAM tightly coupled with raw Global Navigation Satellite System (GNSS) measurements, a first attempt of this kind in the literature to our knowledge. More specifically, reprojection error, IMU pre-integration error and raw GNSS measurement error are jointly minimized within a sliding window, in which the asynchronism between images and raw GNSS measurements is accounted for. In addition, issues such as marginalization, noisy measurements removal, as well as tackling vulnerable situations are also addressed. Experimental results on public dataset in complex urban scenes show that our proposed approach outperforms state-of-the-art visual-inertial SLAM, GNSS single point positioning, as well as a loose coupling approach, including scenes mainly containing low-rise buildings and those containing urban canyons.

Motivation & Objective

• To address the limitations of loose-coupling and EKF-based fusion in visual-inertial SLAM by enabling tight integration of raw GNSS measurements.
• To overcome drift and global frame ambiguity in standard SLAM by fusing raw GNSS pseudorange and Doppler shift with visual and inertial data.
• To improve positioning accuracy in challenging urban environments, including urban canyons and low-texture scenes, where GNSS and vision alone fail.
• To handle sensor asynchrony between cameras, IMUs, and GNSS receivers through precise time alignment in the optimization framework.
• To enhance robustness by integrating noisy measurement removal and marginalization techniques within the optimization pipeline.

Proposed method

• Jointly minimizes reprojection error from visual features, IMU preintegration error, and raw GNSS measurement error (pseudorange and Doppler shift) within a sliding window optimization framework.
• Accounts for asynchronism between image captures, IMU measurements, and GNSS measurements by modeling their temporal offsets explicitly in the optimization.
• Uses a two-step GNSS measurement removal strategy: first a GNSS solver to identify outliers, then a mixed solver combining geometric and kinematic constraints for improved robustness.
• Performs GNSS-SLAM initialization by aligning the VI-SLAM trajectory with GNSS single-point positioning results to estimate initial ECEF-to-ENU and ENU-to-local-world frame transformations.
• Applies marginalization to maintain computational efficiency by removing older states while preserving information through Schur complement techniques.
• Employs a rigid body transformation for trajectory alignment during evaluation to ensure fair comparison with ground truth and baseline methods.

Experimental results

Research questions

• RQ1 Can tight coupling of raw GNSS measurements (pseudorange and Doppler shift) with visual and inertial data improve global localization accuracy in urban environments?
• RQ2 How does joint optimization of visual, inertial, and raw GNSS errors compare to loose-coupling or EKF-based fusion in terms of trajectory accuracy and robustness?
• RQ3 To what extent can sensor asynchrony between cameras, IMUs, and GNSS receivers be mitigated within an optimization framework?
• RQ4 Does the proposed two-stage noisy GNSS measurement removal method improve accuracy without significantly increasing computational cost?
• RQ5 How does the system perform in challenging scenarios such as urban canyons and low-texture environments?

Key findings

• The proposed method achieves a translation RMSE of 14.33 m on the CA 20190828173350 sequence, significantly lower than VINS-mono (19.08 m), GNSS SPP (20.28 m), and VINS-Fusion (14.337 m), demonstrating superior accuracy.
• On the CA 20190828190411 sequence, the method reduces RMSE to 4.455 m using the mixed solver, compared to 5.941 m with the GNSS solver alone, showing the effectiveness of the two-stage removal approach.
• Trajectory completeness is improved across all sequences, with the proposed method maintaining high position availability even in GNSS-degraded environments.
• The system achieves globally consistent trajectories, as confirmed by projection onto Google Earth, showing alignment with real-world paths in complex urban scenes.
• The second noisy measurement removal method (mixed solver) reduces ATE by 25% on average compared to the GNSS solver, albeit with a 25x increase in mean time consumption (24.66 ms vs. 0.389 ms).
• The method outperforms state-of-the-art VI-SLAM (VINS-mono), GNSS single-point positioning (RTKLIB), and loose-coupling (VINS-Fusion) in both translation and rotation accuracy across all tested sequences.