GPR-GSLAM: Gaussian Process Regression-Enhanced Real-Time RGB-D SLAM Using Gaussian Splatting

Abstract

3D Gaussian Splatting (3DGS) has recently revolutionized novel view synthesis and provided a new paradigm for photorealistic Simultaneous Localization and Mapping (SLAM). However, current 3DGS-based RGB-D SLAM systems still face three key challenges: incomplete depth observations due to sensor noise and occlusions, unreliable pose initialization in challenging environments, and the difficulty of balancing reconstruction accuracy with computational efficiency. To address these challenges, we propose Gaussian Process Regression–enhanced Gaussian Splatting SLAM (GPR-GSLAM), a real-time RGB-D SLAM system built around two key components. First, we employ Gaussian Process Regression (GPR) to complete missing depths and introduce an adaptive learning-rate mask to accelerate the optimization of Gaussians initialized from GPR-completed depth while preserving convergence stability. Second, we develop a robust tracking-and-mapping pipeline that combines a Fractional-Weighted Generalized Iterative Closest Point (FW-GICP) with dual-mode tracking for reliable pose initialization, and leverages differentiable 3D Gaussian Splatting for a two-stage pose optimization: a joint SE(3) corrective update of the camera pose and keyframe-initialized Gaussians that rapidly absorbs the residual error from the tracking process, followed by decoupled refinement of camera pose and Gaussian parameters, enabling adaptive map correction and fine-grained accuracy recovery. Quantitative evaluations on TUM RGB-D, Bonn, and Replica show that GPR-GSLAM maintains real-time operation (30 FPS) while achieving strong photorealistic mapping quality and pose accuracy, offering a favorable accuracy-efficiency trade-off.