动态环境下基于语义分割和LK光流法的视觉SLAM算法研究

陈千禧

文章摘要

动态环境下基于语义分割和LK光流法的视觉SLAM算法研究

Research on visual SLAM algorithm based on semantic segmentation and LK optical flow in dynamic environment

投稿时间：2025-05-17 修订日期：2025-06-17

DOI：

英文关键词: simultaneous localization and mapping dynamic feature points semantic segmentation Lucas–Kanade optical flow ghosting elimination

基金项目:四川省智慧旅游研究基地(编号:ZHYJ24-04)资助

作者	单位	邮编
陈千禧^*	四川轻化工大学计算机科学与工程学院	644000

摘要点击次数: 123

全文下载次数: 0

中文摘要:

针对视觉SLAM算法在动态环境下，动态特征点导致定位精度降低、构建的稠密点云地图存在残影的问题，基于语义分割和LK光流法进行动态视觉SLAM算法研究。在RTAB-Map算法中引入基于YOLO11n-Seg的动态特征点剔除模块，获取动态物体的掩膜，剔除动态特征点；针对动态特征点剔除模块的漏检情况，基于动态特征点的光流速度与静态特征点的光流速度不一致，剔除动态特征点，使用静态特征点进行位姿估计；构建稠密点云地图时，利用动态物体掩膜消除残影。实验结果表明：和原RTAB-Map算法相比，改进算法在TUM RGB-D数据集的高动态序列的绝对轨迹误差的RMSE和S.D.分别平均降低77.61%和72.58%；相对位姿误差的平移漂移RMSE和S.D.分别降低71.18%和74.80%，旋转漂移相应指标降低51.06%和51.79%；稠密点云地图残影消除效果明显。

英文摘要:

In dynamic environments, visual SLAM algorithms often suffer from reduced localization accuracy and ghosting artifacts in dense point cloud maps due to dynamic feature points. This study proposes an enhanced dynamic visual SLAM algorithm that integrates semantic segmentation and the Lucas-Kanade (LK) optical flow method. A dynamic feature point removal module based on YOLOv11n-Seg is incorporated into the RTAB-Map framework to obtain dynamic object masks and eliminate dynamic feature points. To address the missed detections of the segmentation module, feature points are further filtered by comparing the optical flow velocities of dynamic and static features, ensuring that only static features are used for pose estimation. During dense map reconstruction, dynamic object masks are utilized to remove ghosting artifacts caused by moving objects. Experimental results on high-dynamic sequences from the TUM RGB-D dataset demonstrate that the proposed method significantly outperforms the original RTAB-Map. Specifically, the Root Mean Square Error (RMSE) and standard deviation (S.D.) of absolute trajectory error are reduced by an average of 77.61% and 72.58%, respectively. The RMSE and S.D. of translational drift in relative pose error decrease by 71.18% and 74.80%, while rotational drift metrics are reduced by 51.06% and 51.79%, respectively. Moreover, the proposed method effectively eliminates ghosting in dense point cloud maps.

View Fulltext 查看/发表评论下载PDF阅读器

关闭