Scholarlink

<p>Simultaneous Localization and Mapping (SLAM) is currently a prominent research topic in robotics, playing a crucial role in the domains of localization and navigation within unfamiliar environments. Traditional SLAM algorithms typically assume static surroundings, yet dynamic elements in real-world environments significantly impact both localization accuracy and robustness, thereby limiting algorithm applicability. To address the challenge of degraded localization accuracy caused by dynamic objects in indoor settings, we propose a dynamic point-line SLAM algorithm based on ORB-SLAM2. To address the issue of insufficient feature points caused by the removal of dynamic points in ORB-SLAM2, we augment the point features with additional line features to introduce more constraints. Additionally, we propose a novel lightweight object detection algorithm called YOLOv5s-Lightweight(YOLOv5sL) based on YOLOv5s. Experimental results on the VOC dataset demonstrate that our model achieves better real-time performance while maintaining high accuracy compared to the original YOLOv5s. Furthermore, leveraging the RGB-D model of ORB-SLAM2, we present an adaptive epipolar constraint-based dynamic feature filtering algorithm. This algorithm effectively removes dynamic points and lines by utilizing prior information from object detection networks and incorporating depth information through a new tracking thread. As a result, it efficiently filters out dynamic feature points and significantly reduces the influence of dynamic objects. Finally, we validate the effectiveness of our improved algorithm using dynamic sequences from the TUM RGB-D dataset, showing notable reductions in Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), thus confirming its efficacy.</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

Dynamic point-line SLAM based on lightweight object detection

  • 期刊论文

  • 发表日期：

    2025年1月3日

查看全文
  • Jiaqi Zhong,
  • Huaming Qian

摘要

Simultaneous Localization and Mapping (SLAM) is currently a prominent research topic in robotics, playing a crucial role in the domains of localization and navigation within unfamiliar environments. Traditional SLAM algorithms typically assume static surroundings, yet dynamic elements in real-world environments significantly impact both localization accuracy and robustness, thereby limiting algorithm applicability. To address the challenge of degraded localization accuracy caused by dynamic objects in indoor settings, we propose a dynamic point-line SLAM algorithm based on ORB-SLAM2. To address the issue of insufficient feature points caused by the removal of dynamic points in ORB-SLAM2, we augment the point features with additional line features to introduce more constraints. Additionally, we propose a novel lightweight object detection algorithm called YOLOv5s-Lightweight(YOLOv5sL) based on YOLOv5s. Experimental results on the VOC dataset demonstrate that our model achieves better real-time performance while maintaining high accuracy compared to the original YOLOv5s. Furthermore, leveraging the RGB-D model of ORB-SLAM2, we present an adaptive epipolar constraint-based dynamic feature filtering algorithm. This algorithm effectively removes dynamic points and lines by utilizing prior information from object detection networks and incorporating depth information through a new tracking thread. As a result, it efficiently filters out dynamic feature points and significantly reduces the influence of dynamic objects. Finally, we validate the effectiveness of our improved algorithm using dynamic sequences from the TUM RGB-D dataset, showing notable reductions in Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), thus confirming its efficacy.