A Resilient Method for Visual-Inertial Fusion Based on Covariance Tuning

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Accurate localization and pose estimation are critical for visual-inertial simultaneous localization and mapping (viSLAM) systems, especially in challenging environments.
• Optimizing the fusion of visual and inertial data through weight tuning is essential for enhancing viSLAM performance.

Purpose of the Study:

• To develop a robust viSLAM algorithm that improves localization and pose precision by adaptively tuning sensor fusion weights.
• To enhance the performance of viSLAM systems without relying on traditional closed-loop detection mechanisms.

Main Methods:

• A novel resilient viSLAM algorithm based on covariance tuning is proposed.
• The method computes unit-weight root-mean-square error (RMSE) for visual reprojection and IMU preintegration during back-end optimization.
• A covariance tuning function is constructed to generate an updated covariance matrix for subsequent nonlinear optimization.

Main Results:

• The proposed covariance tuning approach significantly improved pose and localization precision in viSLAM.
• The algorithm demonstrated superior performance compared to established open-source frameworks like OKVIS, R-VIO, and VINS-Mono.
• Performance enhancements were validated across all difficulty levels on the EuRoc dataset.

Conclusions:

• Covariance tuning offers an effective strategy for improving viSLAM accuracy without closed-loop detection.
• The developed resilient viSLAM algorithm presents a significant advancement in precise localization and pose estimation for complex scenarios.