Scholarlink

<p>To further enhance the adaptability of the SLAM system to dynamic environments in reality and improve the pose estimation accuracy of the SLAM system, this paper proposes a dynamic environment DG-SLAM based on YOLOv5 target detection, combined with multi-scale feature fusion and geometric constraints. The tracked feature points are divided into dynamic, static, and hidden dynamic points by the model. The dynamic targets in the environment are accurately identified by combining the target detection algorithm with multi-scale feature fusion, and then the dynamic feature points are eliminated; based on geometric constraints, static points and latent dynamic points are distinguished by setting thresholds, constructing a hierarchical semantic–geometric joint constraint dynamic scene robust visual SLAM framework, which effectively reduces the impact of dynamic targets on system performance. The proposed method was experimentally evaluated on challenging dynamic sequences in the TUM and Bonn datasets. The results show that, in both CPU environments and TUM dynamic dataset sequences, the absolute trajectory error (ATE) of DG-SLAM is significantly reduced compared to ORB-SLAM3 and YOLO-Fastest-SLAM, which perform well in dynamic environments. Furthermore, it improves the stability of pose estimation in rotational RPE. Although it slightly decreases in translational RPE, the overall trajectory still maintains good consistency. At the same time, the trajectory accuracy is good in the Bonn dataset, indicating that the proposed method can achieve good accuracy and robustness in dynamic environments.</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

DG-SLAM: research on dynamic feature point removal of visual SLAM by integrating multi-scale YOLOv5 and geometric constraints

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月31日

查看全文
  • Bo Liu,
  • Juwei Zhang,
  • Bingyi Ren,
  • Xuguang Hu,
  • Tong Wang,
  • Yuxuan Liu

摘要

To further enhance the adaptability of the SLAM system to dynamic environments in reality and improve the pose estimation accuracy of the SLAM system, this paper proposes a dynamic environment DG-SLAM based on YOLOv5 target detection, combined with multi-scale feature fusion and geometric constraints. The tracked feature points are divided into dynamic, static, and hidden dynamic points by the model. The dynamic targets in the environment are accurately identified by combining the target detection algorithm with multi-scale feature fusion, and then the dynamic feature points are eliminated; based on geometric constraints, static points and latent dynamic points are distinguished by setting thresholds, constructing a hierarchical semantic–geometric joint constraint dynamic scene robust visual SLAM framework, which effectively reduces the impact of dynamic targets on system performance. The proposed method was experimentally evaluated on challenging dynamic sequences in the TUM and Bonn datasets. The results show that, in both CPU environments and TUM dynamic dataset sequences, the absolute trajectory error (ATE) of DG-SLAM is significantly reduced compared to ORB-SLAM3 and YOLO-Fastest-SLAM, which perform well in dynamic environments. Furthermore, it improves the stability of pose estimation in rotational RPE. Although it slightly decreases in translational RPE, the overall trajectory still maintains good consistency. At the same time, the trajectory accuracy is good in the Bonn dataset, indicating that the proposed method can achieve good accuracy and robustness in dynamic environments.