Scholarlink

<p>A&#xa0;computationally efficient method for enhancing the robustness of monocular simultaneous localization and mapping (SLAM) systems based on the ORB-SLAM3 architecture for operation in dynamic environments is presented. The negative impact of moving objects on epipolar geometry estimation and global scale drift is mathematically formalized. A&#xa0;method for proactive video stream filtering using convolutional neural networks (YOLO-Seg) is proposed, where dynamic objects are identified and excluded from the feature extraction process using binary masks. The necessity of applying morphological mask dilation to compensate for the FAST detector’s aperture is strictly justified. Experiments on the KITTI dataset demonstrate a&#xa0;reduction in the Absolute Trajectory Error (ATE) by 23–31% and an improvement in Relative Pose Error (RPE) metrics. Implementation as a&#xa0;ROS2 node graph proves the applicability of the method in real-time control loops (up to 30&#xa0;frames per second) without modifying the mathematical core of the SLAM optimizer.</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

Enhancing the robustness of monocular SLAM in dynamic scenes based on semantic image filtering

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Pavel Olegovich Yaroschuk,
  • Marina Vasilievna Polovinkina,
  • Igor Petrovich Polovinkin

摘要

A computationally efficient method for enhancing the robustness of monocular simultaneous localization and mapping (SLAM) systems based on the ORB-SLAM3 architecture for operation in dynamic environments is presented. The negative impact of moving objects on epipolar geometry estimation and global scale drift is mathematically formalized. A method for proactive video stream filtering using convolutional neural networks (YOLO-Seg) is proposed, where dynamic objects are identified and excluded from the feature extraction process using binary masks. The necessity of applying morphological mask dilation to compensate for the FAST detector’s aperture is strictly justified. Experiments on the KITTI dataset demonstrate a reduction in the Absolute Trajectory Error (ATE) by 23–31% and an improvement in Relative Pose Error (RPE) metrics. Implementation as a ROS2 node graph proves the applicability of the method in real-time control loops (up to 30 frames per second) without modifying the mathematical core of the SLAM optimizer.