Scholarlink

<p>The random appearance of dynamic targets in the field-of-view (FoV) of mobile robot can seriously interfere with the sensor measurement of its SLAM system and may even render the system ineffective. Along these lines, we propose a visual SLAM (vSLAM) scheme based on RGB-D measurement and compact 3D Gaussian environment representation. First, YOLOv11 is introduced to perform instance segmentation on the input RGB images and extract the prior dynamic masks; meanwhile, it is fused with potential dynamic masks identified through optical flow and epipolar geometry to obtain a combined dynamic mask, which are subsequently optimized by using the residual entropy and depth consistency to accurately reject the dynamic features. Then, to simultaneously balance the stability of pose tracking and the efficiency of mapping, a two-stage keyframe screening mechanism is constructed that combines “localization-oriented” and “mapping-oriented” approaches. On this basis, a compact 3D Gaussian splatting (3DGS) technique is further introduced to render and construct the scene map. For each 3D Gaussian ellipsoid, Kalman filtering (KF) is utilized to estimate the weight of its volume and opacity, and the contribution weight of its multi-view is calculated, and then a self-adaption Markov model (SAMM) based weighted average method is developed to obtain the final weight and perform pruning. A series of simulations and experimental tests in real scene show that our algorithm can accurately remove dynamic components from RGB-D measurements, providing accurate localization results and efficiently generating realistically rendered maps.</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

Approach to visual SLAM based on RGB-D measurement and compact 3D Gaussian environment representation in dynamic scenes

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月23日

查看全文
  • Enbo Zhang,
  • Tian Wang,
  • Hao Liu,
  • Xuanzhen Chen,
  • Jingwen Luo

摘要

The random appearance of dynamic targets in the field-of-view (FoV) of mobile robot can seriously interfere with the sensor measurement of its SLAM system and may even render the system ineffective. Along these lines, we propose a visual SLAM (vSLAM) scheme based on RGB-D measurement and compact 3D Gaussian environment representation. First, YOLOv11 is introduced to perform instance segmentation on the input RGB images and extract the prior dynamic masks; meanwhile, it is fused with potential dynamic masks identified through optical flow and epipolar geometry to obtain a combined dynamic mask, which are subsequently optimized by using the residual entropy and depth consistency to accurately reject the dynamic features. Then, to simultaneously balance the stability of pose tracking and the efficiency of mapping, a two-stage keyframe screening mechanism is constructed that combines “localization-oriented” and “mapping-oriented” approaches. On this basis, a compact 3D Gaussian splatting (3DGS) technique is further introduced to render and construct the scene map. For each 3D Gaussian ellipsoid, Kalman filtering (KF) is utilized to estimate the weight of its volume and opacity, and the contribution weight of its multi-view is calculated, and then a self-adaption Markov model (SAMM) based weighted average method is developed to obtain the final weight and perform pruning. A series of simulations and experimental tests in real scene show that our algorithm can accurately remove dynamic components from RGB-D measurements, providing accurate localization results and efficiently generating realistically rendered maps.