Scholarlink

<p>We present a novel RGB-D SLAM method that employs 3D Gaussian splatting (3DGS) as the core geometric representation for compact and efficient scene modeling. To address the accuracy–real-time trade-off in conventional camera tracking methods, which typically rely on minimizing photometric and geometric losses, we propose a camera pose estimation method that integrates surface normal constraints into a multi-level ICP algorithm, thereby improving both robustness and computational efficiency. Furthermore, we introduce a dynamic keyframe selection strategy that jointly considers Gaussian overlap coefficient and camera motion to effectively reduce redundant computations. For high-fidelity 3D reconstruction, we develop a sliding window-based joint optimization method for Gaussian parameters, supplemented by a multi-view densification strategy to improve scene completeness and geometric consistency. Extensive experiments demonstrate that our method outperforms existing approaches in pose estimation accuracy while achieving more photorealistic novel view synthesis. Code is available at: <a href="https://github.com/Justzhongjj/MDGS_SLAM">https://github.com/Justzhongjj/MDGS_SLAM</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MDGS-SLAM: real-time RGB-D Gaussian-SLAM with multi-view densification for scene reconstruction

  • 期刊论文

  • 发表日期：

    2026年1月19日

查看全文
  • Junjin Zhong,
  • Jinlong Shi,
  • Qiang Qian,
  • SuQin Bai,
  • Wei Teng,
  • Linbin Pang

摘要

We present a novel RGB-D SLAM method that employs 3D Gaussian splatting (3DGS) as the core geometric representation for compact and efficient scene modeling. To address the accuracy–real-time trade-off in conventional camera tracking methods, which typically rely on minimizing photometric and geometric losses, we propose a camera pose estimation method that integrates surface normal constraints into a multi-level ICP algorithm, thereby improving both robustness and computational efficiency. Furthermore, we introduce a dynamic keyframe selection strategy that jointly considers Gaussian overlap coefficient and camera motion to effectively reduce redundant computations. For high-fidelity 3D reconstruction, we develop a sliding window-based joint optimization method for Gaussian parameters, supplemented by a multi-view densification strategy to improve scene completeness and geometric consistency. Extensive experiments demonstrate that our method outperforms existing approaches in pose estimation accuracy while achieving more photorealistic novel view synthesis. Code is available at: https://github.com/Justzhongjj/MDGS_SLAM.