Visual SLAM with 3D Gaussian Splatting and Generalized ICP

Simultaneous localization and mapping (SLAM) plays a crucial role in robotics, augmented reality (AR), and virtual reality (VR). Recently, SplaTAM, based on 3D Gaussian Splatting, has significantly improved dense mapping and camera pose estimation efficiency. However, it remains susceptible to pose drift in texture-poor or highly dynamic scenes, limiting its broader applicability. To address this issue, this paper proposes a novel RGB-D SLAM framework integrating 3D Gaussian Splatting and Generalized ICP (GICP). Specifically, our method initially employs 3D Gaussian Splatting to perform coarse pose estimation based on photometric and depth information. Subsequently, we apply the GICP method with a plane-to-plane error model for precise point cloud registration, further enhancing the robustness and accuracy of pose estimation. Extensive experiments conducted on the public Replica dataset demonstrate that our approach outperforms the original SplaTAM method, highlighting its significant advantages for precise and robust mapping in diverse environments.