Globally Consistent Submap-Based 3D Gaussian SLAM with Gradient-Guided Tracking and Rendering-Aware Loop Verification

Abstract

Dense RGB-D SLAM aims to estimate camera trajectories while reconstructing high-quality three-dimensional scenes from sequential RGB-D frames. Recent SLAM methods based on 3D Gaussian Splatting enable efficient differentiable rendering and high-fidelity mapping. However, three limitations remain in long RGB-D sequences: uniform pixel weighting treats image regions with different contributions to pose estimation equally, preventing regions that provide stronger pose constraints from being sufficiently emphasized and reducing tracking robustness; loop closure candidates are verified at the registration level without an explicit rendering-consistency check against the observed scene geometry, potentially admitting geometrically inconsistent loop closures and introducing erroneous constraints into the pose graph; and pose graph optimization corrects submap poses without jointly adapting the corresponding Gaussian parameters to the corrected trajectory, resulting in inconsistencies between the corrected camera trajectory and the reconstructed Gaussian map. To address these limitations, a loop-aware submap-based 3DGS RGB-D SLAM method is proposed in this paper as an extension of the LoopSplat framework, and three dedicated components are developed to address the aforementioned limitations. During front-end tracking, a gradient-guided RGB-D tracking strategy is proposed, in which higher optimization weights are assigned to image regions that provide stronger constraints for pose estimation, so that their contributions are sufficiently emphasized and tracking robustness is improved. Before loop constraints are added to the pose graph, a rendering-consistency-based loop verification module is developed, through which candidate loop closures are evaluated using depth rendering residuals, while those whose estimated relative poses are inconsistent with the observed scene geometry are rejected, thereby preventing erroneous constraints from being introduced into the pose graph. After pose graph optimization, a loop-aware Gaussian Bundle Adjustment scheme is formulated to refine loop-related submap poses and the corresponding Gaussian parameters in two sequential stages, followed by post-PGO Gaussian-parameter refinement to further reduce residual map inconsistency and improve the consistency between the corrected camera trajectory and the reconstructed Gaussian map. Results across the eight Replica scenes show that, under identical experimental settings, the proposed method achieves higher trajectory accuracy and RGB rendering quality than the reproduced LoopSplat baseline. Specifically, the average aligned ATE RMSE is reduced from 0.58 cm to 0.52 cm, the average PSNR is increased from 35.68 dB to 36.27 dB, the average SSIM is increased from 0.9807 to 0.9832, and the average LPIPS is reduced from 0.1223 to 0.1173.