Scholarlink

<p>Dynamic objects, low-texture regions, and motion blur remain major challenges for robust RGB-D visual simultaneous localization and mapping. This paper presents GDN-SLAM, a geometry-guided dynamic SLAM framework that integrates point-line feature consistency, dual-stage dynamic feature suppression, and object-level neural scene constraints. Built upon ORB-SLAM2, GDN-SLAM first enhances feature reliability through image deblurring and joint point-line extraction. A hierarchical geometric consistency model then couples point-line projection, orientation, scale, and global pose constraints to improve feature association in weakly textured and dynamic regions. To reduce dynamic interference, we design an explicit–implicit optimization strategy that combines epipolar geometry and YOLO-based object detection for dynamic feature removal, followed by motion-aware adaptive weighting for residual dynamic features. In addition, a lightweight object-level NeRF module provides geometric, semantic, and photometric soft constraints for pose refinement at selected keyframes. Experiments on the TUM RGB-D and Bonn RGB-D dynamic datasets show that GDN-SLAM improves localization accuracy and robustness over representative traditional and dynamic SLAM baselines, while maintaining practical runtime efficiency. The source code and experimental materials are available at <a href="https://github.com/JunjieHuang-2001/GDN-SLAM">https://github.com/JunjieHuang-2001/GDN-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

GDN-SLAM: geometry-guided dynamic RGB-D SLAM with point-line consistency and object-level neural scene constraints

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Huilin Liu,
  • Junjie Huang,
  • Lunqi Yu,
  • Chang Su,
  • Wanqi Ma

摘要

Dynamic objects, low-texture regions, and motion blur remain major challenges for robust RGB-D visual simultaneous localization and mapping. This paper presents GDN-SLAM, a geometry-guided dynamic SLAM framework that integrates point-line feature consistency, dual-stage dynamic feature suppression, and object-level neural scene constraints. Built upon ORB-SLAM2, GDN-SLAM first enhances feature reliability through image deblurring and joint point-line extraction. A hierarchical geometric consistency model then couples point-line projection, orientation, scale, and global pose constraints to improve feature association in weakly textured and dynamic regions. To reduce dynamic interference, we design an explicit–implicit optimization strategy that combines epipolar geometry and YOLO-based object detection for dynamic feature removal, followed by motion-aware adaptive weighting for residual dynamic features. In addition, a lightweight object-level NeRF module provides geometric, semantic, and photometric soft constraints for pose refinement at selected keyframes. Experiments on the TUM RGB-D and Bonn RGB-D dynamic datasets show that GDN-SLAM improves localization accuracy and robustness over representative traditional and dynamic SLAM baselines, while maintaining practical runtime efficiency. The source code and experimental materials are available at https://github.com/JunjieHuang-2001/GDN-SLAM.