Abstract:
Recent SLAM systems have adopted 3D Gaussian Splatting (3DGS) to generate photorealistic maps. Despite its high fidelity, scaling 3DGS-SLAM to large scenes remains challenging. In most implementations, 3DGS maintains a single globally shared set of Gaussians, which can cause catastrophic forgetting and excessive GPU memory growth under continual online optimization. Submap partitioning alleviates these issues, but heuristic split rules often trigger unnecessary splits, leading to unstable initialization and inconsistent mapping. We propose a novel 3DGS-based SLAM framework that integrates efficient submap partitioning (ESP) with a Background Submap Management (BSM) module. ESP identifies reliable split moments by monitoring Gaussian representation quality and filtering out unstable observations that may compromise mapping stability. Meanwhile, BSM refines Gaussians in the background and manages memory to support scalable deployment. Experiments on benchmark datasets demonstrate robust SLAM performance while reducing peak GPU memory usage by up to 48%. Ablation studies further validate the contribution of each component and the overall robustness of the system.