基于语义分割的动态VSLAM建图线程实现OA
Realization of dynamic V-SLAM mapping thread based on semantic segmentation
针对动态环境下传统视觉即时定位与地图构建(visual simultaneous localization and mapping,VSLAM)建图存在信息冗余与地图拖影的问题,在ORB-SLAM3 框架下,本文设计了基于 DeepLabV3+网络的动态 VSLAM 建图线程.首先,通过轻量级 DeepLabV3+网络对建图关键帧进行语义分割,实现剔除 VSLAM 地图关键帧中的动态区域;然后,将剔除动态特征后的位姿数据与深度学习信息相结合,构建了基于 MobileNetV3 的轻量级DeepLabV3+网络,生成排除动态特征的三维静态稠密点云地图及八叉树.实验结果表明,本文所设计方法有效减少了动态环境下 VSLAM 地图信息冗余量和地图拖影,降低了地图空间占用比.
In order to solve the problems of redundant information and ghosting in map construction of traditional VSLAM in dynamic environments,a dynamic VSLAM mapping thread method based on DeepLabV3+is designed and ap-plied on ORB-SLAM3 architecture.Firstly,Key mapping frames are semantically segmented with a lightweight DeepLabV3+network to exclude dynamic regions.Secondly,the pose data after eliminating dynamic features is integrated with deep learning information,and then the point cloud map and octree map with semantic information are generated.Finally,the experimental results demonstrate that the maps constructed by the improved mapping method are with less redundant information and fewer ghosts which decrease the map space occupancy.
黄卫华;张泽宇;章政
武汉城市学院机电工程学部 武汉 430083||武汉科技大学人工智能与自动化学院 武汉 430081武汉科技大学人工智能与自动化学院 武汉 430081武汉科技大学人工智能与自动化学院 武汉 430081
视觉即时定位与地图构建语义分割DeepLabV3+网络动态环境
visual simultaneous localization and mappingsemantic segmentationDeepLabV3+networkdynamic environment
《高技术通讯》 2026 (6)
586-592,7
国家自然科学基金(62303359)资助项目.
评论