Uncertainty Modeling in Gaussian Splatting for RGB-D Simultaneous Localization and Mapping

This thesis presents an uncertainty-aware, photorealistic SLAM framework based on 3D Gaussian Splatting. Conventional SLAM and Gaussian-based mapping systems often face instability under dynamic lighting, sensor noise, and scene variations, as they do not account for the propagation of uncertainty across geometric, photometric, and temporal domains. To address this limitation, this work incorporates uncertainty modeling in Gaussian-splat representations and develops a confidence-aware SLAM pipeline. We first introduce a dual-depth fusion mechanism that integrates RGB-D sensor measurements with Gaussian-rendered depth, each weighted by per-pixel variance derived from a calibrated noise model. Second, a temporal uncertainty graph is formulated to monitor merge and split events among Gaussians, serving as an indicator of local scene consistency. Lastly, a new uncertainty-aware loss function combines photometric, geometric, and graph-regularization terms, with KL-variance control for stable Gaussian updates. The resulting system extends existing GS-based SLAM methods with uncertainty decomposition across geometric, photometric, and temporal domains. Experiments on Replica RGB-D and TUM RGB-D datasets demonstrate improved stability and robustness of photorealistic SLAM. The proposed method provides a unified uncertainty-aware formulation for photorealistic SLAM, connecting Gaussian Splatting and probabilistic visual mapping.