Scholarlink

<p>In semi-structured environments characterized by static planes and dynamic objects, feature-based SLAM systems often suffer from performance degradation. To address this, we propose a semantic SLAM framework that integrates advanced object detection with plane geometry constraints. Our system leverages the YOLOv7 algorithm for object recognition and real-time point cloud normal vector clustering to extract planes from depth images. Furthermore, we propose an algorithm for eliminating dynamic feature points based on the spatial geometric relationships between objects and planes, enhancing the robustness of localization and mapping processes. To reduce the propagation of cumulative errors, we decouple pose estimation into rotational and translational components, and then use surface normal vectors to compute the rotation estimation within the Atlanta World framework. Moreover, the backend employs factor graph optimization incorporating geometric constraints, semantic information, and data association metrics. Comprehensive evaluations conducted on the TUM RGB-D benchmark datasets and real-world scenarios involving intelligent manufacturing cell demonstrate that our method achieves improvements in trajectory accuracy compared to state-of-the-art approaches.</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

Semantic SLAM for dynamic semi-structured environments using object and plane detection

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Shiqi Li,
  • Yibin Liu,
  • Lequn Fu,
  • Meicheng Han,
  • Youjun Xiong

摘要

In semi-structured environments characterized by static planes and dynamic objects, feature-based SLAM systems often suffer from performance degradation. To address this, we propose a semantic SLAM framework that integrates advanced object detection with plane geometry constraints. Our system leverages the YOLOv7 algorithm for object recognition and real-time point cloud normal vector clustering to extract planes from depth images. Furthermore, we propose an algorithm for eliminating dynamic feature points based on the spatial geometric relationships between objects and planes, enhancing the robustness of localization and mapping processes. To reduce the propagation of cumulative errors, we decouple pose estimation into rotational and translational components, and then use surface normal vectors to compute the rotation estimation within the Atlanta World framework. Moreover, the backend employs factor graph optimization incorporating geometric constraints, semantic information, and data association metrics. Comprehensive evaluations conducted on the TUM RGB-D benchmark datasets and real-world scenarios involving intelligent manufacturing cell demonstrate that our method achieves improvements in trajectory accuracy compared to state-of-the-art approaches.