Abstract
Back-end optimization is a key process to eliminate the cumulative error in Visual Simultaneous Localization and Mapping (VSLAM). Existing VSLAM frameworks often use kernel function-based back-end optimization methods. However, these methods typically rely on fixed kernel parameters based on the chi-square test, assuming Gaussian-distributed reprojection errors. In practice, though, reprojection errors are not always Gaussian, which can reduce robustness and accuracy. Therefore, we propose a data-adaptive iteratively reweighted robust kernel (DA-IRRK) approach, which combines median absolute deviation (MAD) with iteratively reweighted strategies. The robustness parameters are adaptively adjusted according to the MAD of reprojection errors, and the Huber kernel function is used to demonstrate the implementation of the back-end optimization process. The method is compared with other robust function-based approaches via the EuRoC dataset and the KITTI dataset, showing adaptability across different VSLAM frameworks and demonstrating significant improvements in trajectory accuracy on the vast majority of dataset sequences. The statistical analysis of the results from the perspective of reprojection error indicates DA-IRRK can tackle non-Gaussian noises better than the compared methods.