Pre-Dog-Leg: A Feature Optimization Method for Visual Inertial SLAM Based on Adaptive Preconditions.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Monocular visual-inertial SLAM faces challenges due to ill-posed Hessian matrices and unobservable feature point depth.
• This leads to convergence difficulties and reduced system robustness in localization and mapping.

Purpose of the Study:

• To propose a Pre-Dog-Leg feature optimization method with an adaptive preconditioner to address Hessian matrix ill-posedness in monocular visual-inertial SLAM.
• To improve convergence speed, robustness, and localization accuracy.

Main Methods:

• A multi-candidate initialization method with robust characteristics using multiple depth assumptions and geometric consistency constraints.
• Construction of a hybrid SPAI-Jacobi adaptive preconditioner to identify and address Hessian matrix pathology.
• Development of a hybrid adaptive preconditioner for the Dog-Leg method, mapping problems to a well-conditioned space while maintaining optimization equivalence.

Main Results:

• Reduced Hessian matrix condition numbers by a factor of 7.9.
• Effective suppression of outliers and significant improvement in overall convergence time.
• Reduced absolute trajectory error by up to 16.48% (vs. RVIO2), 20.83% (vs. VINS-mono), and 34.0% (vs. OpenVINS on dynamic sequences).

Conclusions:

• The proposed Pre-Dog-Leg optimization method with an adaptive preconditioner significantly enhances the performance of monocular visual-inertial SLAM.
• The algorithm demonstrates higher localization accuracy and stronger system robustness, particularly in challenging, dynamic environments.