DA-IRRK: Data-Adaptive Iteratively Reweighted Robust Kernel-Based Approach for Back-End Optimization in Visual SLAM.

Area of Science:

• Robotics
• Computer Vision
• Simultaneous Localization and Mapping (SLAM)

Background:

• Back-end optimization is crucial for minimizing cumulative errors in Visual Simultaneous Localization and Mapping (VSLAM).
• Current VSLAM frameworks often employ kernel function-based optimization, assuming Gaussian-distributed reprojection errors.
• This assumption can lead to reduced robustness and accuracy when errors deviate from Gaussian distributions.

Purpose of the Study:

• To develop a novel back-end optimization method for VSLAM that addresses the limitations of fixed kernel parameters.
• To improve the robustness and accuracy of VSLAM systems when dealing with non-Gaussian reprojection errors.
• To introduce a data-adaptive approach for robust kernel optimization in VSLAM.

Main Methods:

• Proposed a data-adaptive iteratively reweighted robust kernel (DA-IRRK) approach.
• Integrated median absolute deviation (MAD) with iteratively reweighted strategies for adaptive parameter adjustment.
• Utilized the Huber kernel function within the DA-IRRK framework for back-end optimization.

Main Results:

• DA-IRRK demonstrated significant improvements in trajectory accuracy across most sequences in the EuRoC and KITTI datasets.
• The proposed method showed adaptability across different VSLAM frameworks.
• Statistical analysis confirmed DA-IRRK's superior ability to handle non-Gaussian noise compared to existing robust methods.

Conclusions:

• The DA-IRRK approach offers a more robust and accurate solution for VSLAM back-end optimization.
• Adaptive adjustment of robustness parameters based on data characteristics is key to handling non-Gaussian errors.
• This method enhances VSLAM performance in real-world scenarios where error distributions may not be strictly Gaussian.