Credibility-driven sampling for random sample consensus-based outlier removal in visual SLAM

Standard RANSAC scales poorly as its computational cost grows rapidly with scene complexity, and GMS-RANSAC still suffers from redundant iterations without breaking the accuracy-efficiency trade-off. To address this, we propose Improved GMS-RANSAC, a novel credibility-driven outlier removal method that pioneers 3×3 grid neighborhood credibility quantification and halving-based grouping. It defines match credibility via 3×3 grid neighborhood matching counts, sorts matches by credibility, and adopts a halving-based grouping strategy to prioritize high-credibility samples for RANSAC, replacing blind random sampling to optimize RANSAC inlier probability and fundamentally cut redundant iterations. Experiments on KITTI, TUM desk, and TUM doll datasets demonstrate that our method achieves 77.02% precision on TUM desk, maintains accuracy comparable to original GMS-RANSAC on other datasets, and reduces average runtime by 34.20%. Furthermore, integrated into ORB-SLAM2, it reduces system runtime by 4.61%, decreases relative pose error by 11.7% and rotational error by 4.9% while preserving accuracy, providing a lightweight, efficient solution for real-time visual SLAM.