59 / 2025-05-14 11:38:51
Dense Point Cloud Map Generation via Attention-Based Monocular-Inertial SLAM
Monocular-Inertial SLAM, Dense Point Mapping, Pose estimation, Feature extraction
全文待审
智昊 谢 / 北京信息科技大学
清华 苏 / 北京信息科技大学
智纬 白 / 北京信息科技大学
立勇 王 / 北京信息科技大学
越 宋 / 北京信息科技大学
毅政 刘 / 北京信息科技大学
Monocular vision technology has attracted considerable attention in visual SLAM research due to its simple structure and low cost. However, it faces inherent limitations in dense map reconstruction, primarily due to the absence of direct depth information. To address this challenge, we propose a dense point cloud reconstruction method based on a monocular-inertial sparse visual SLAM framework. This approach enhances the ORB-SLAM3 system performance by integrating a monocular depth estimation network based on deep learning and attention mechanisms. The predicted dense depth maps are fused with camera poses estimated by SLAM to generate a globally consistent 3D point cloud map. Experimental results conducted on open datasets demonstrate that our method maintains accurate localization performance while producing dense and structurally complete point cloud maps. The depth estimation accuracy on KITTI achieves 90.1% of pixels satisfying the error threshold of δ1< 1.25. Consequently, this system is well suited for large-scale outdoor 3D reconstruction and environmental perception tasks, offering high practical value.

 
重要日期
  • 会议日期

    08月01日

    2025

    08月04日

    2025

  • 06月15日 2025

    初稿截稿日期

主办单位
中国机械工程学会设备智能运维分会
承办单位
新疆大学
移动端
在手机上打开
小程序
打开微信小程序
客服
扫码或点此咨询