Scholarlink

<p>In the context of dynamic scenes, robots face challenges in achieving accurate localization and mapping through vision-based Simultane- ous Localization and Mapping (SLAM) tasks. Moving objects often introduce interference, resulting in decreased localization accuracy and system robustness. This paper proposes a deep learning-based SLAM system, built upon ORB-SLAM2, to address these issues. The pro- posed approach combines the GCNv2 feature extraction algorithm with ORB-SLAM2. Notably, the two primary inter-frame tracking meth- ods in ORB-SLAM2, namely reference keyframe tracking and constant velocity motion model tracking, have been enhanced. To mitigate inter- ference from moving objects, a dynamic point removal algorithm is introduced. This algorithm leverages Yolo-Fastest for detecting dynamic objects in advance and employs Farneback optical flow to identify changes caused by static objects. Additionally, a static scene map con- struction algorithm is developed. This algorithm incorporates a point cloud mapping module to generate dense point cloud maps. Static keyframes are then utilized to construct static scene maps. To evalu- ate the proposed SLAM system’s performance, it is compared against ORB-SLAM2, DS-SLAM, and DynaSLAM using the TUM RGB-D dataset. Real scene testing is conducted as well. The results highlight the superior overall performance of the proposed SLAM system com- pared to ORB-SLAM2, DS-SLAM, and DynaSLAM. Furthermore, it demonstrates its feasibility and effectiveness in real-world scenarios.</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

Deep learning-based visual SLAM for indoor dynamic scenes

  • 期刊论文

  • 发表日期：

    2025年2月11日

查看全文
  • Zhendong Xu,
  • Yong Song,
  • Bao Pang,
  • Qingyang Xu,
  • Xianfeng Yuan

摘要

In the context of dynamic scenes, robots face challenges in achieving accurate localization and mapping through vision-based Simultane- ous Localization and Mapping (SLAM) tasks. Moving objects often introduce interference, resulting in decreased localization accuracy and system robustness. This paper proposes a deep learning-based SLAM system, built upon ORB-SLAM2, to address these issues. The pro- posed approach combines the GCNv2 feature extraction algorithm with ORB-SLAM2. Notably, the two primary inter-frame tracking meth- ods in ORB-SLAM2, namely reference keyframe tracking and constant velocity motion model tracking, have been enhanced. To mitigate inter- ference from moving objects, a dynamic point removal algorithm is introduced. This algorithm leverages Yolo-Fastest for detecting dynamic objects in advance and employs Farneback optical flow to identify changes caused by static objects. Additionally, a static scene map con- struction algorithm is developed. This algorithm incorporates a point cloud mapping module to generate dense point cloud maps. Static keyframes are then utilized to construct static scene maps. To evalu- ate the proposed SLAM system’s performance, it is compared against ORB-SLAM2, DS-SLAM, and DynaSLAM using the TUM RGB-D dataset. Real scene testing is conducted as well. The results highlight the superior overall performance of the proposed SLAM system com- pared to ORB-SLAM2, DS-SLAM, and DynaSLAM. Furthermore, it demonstrates its feasibility and effectiveness in real-world scenarios.