摘要
针对二维图像目标识别准确率高但缺乏空间位置信息,三维点云几何信息精确但无语义信息的特点,为结合不同维度数据的优势,提出一种三维点云与二维图像双模态融合的空间目标部件识别方法。首先基于DeepLabv3+网络进行图像目标分割,并对分割目标进行语义标记;其次,提出一种最近邻查找方法降低重建位姿误差带来的三维点云与二维目标映射过程中的语义信息缺失,实现三维点云与二维图像目标间的准确语义关联;最后,利用全连接条件随机场对具备语义信息的三维点云识别结果进行优化,得到更加精细的点云语义标签,融合多视角数据得到最终点云部件识别结果。仿真结果表明相较于传统聚类分割方法,本文方法可以有效地识别出空间目标的各部件,总体识别精度优于95%。
Aiming at the characteristics of high accuracy in image target recognition but lack of spatial location information,and three-dimensional point clouds have accurate geometric locations without semantic information,a components recognition method for space target based on bi-modal fusion of 3D point cloud and image is proposed to combine the advantages of different dimensional data.Firstly,with the introduction of DeepLabv3+network,the propose method can segment the image to get high-precision pixels’labels of the target.Then we project the 3D point cloud and use nearest neighbor search in target pixels to get each 3D point’s semantic label associated with the target’s pixels to reduce the lack of 3D point semantics caused by pose errors.Finally,we use the fully connected conditional random field(DenseCRF)to optimize the recognition results of 3D point clouds with semantic information to obtain more refined semantic labels.The simulation results have shown that compared with the traditional cluster segmentation,our method can effectively recognize the components of the space target,the overall recognition accuracy is better than 95%.
作者
袁萌萌
张泽旭
YUAN Mengmeng;ZHANG Zexu(School of Astronautics,Harbin Institute of Technology,Harbin 150001,China)
出处
《宇航学报》
EI
CAS
CSCD
北大核心
2023年第5期796-804,共9页
Journal of Astronautics
基金
国家自然科学基金(61573247)
基础加强计划(173计划)(2020-JCJQ-ZD-015-00)
中央高校基本科研业务费专项资金。
作者简介
袁萌萌(1998-),女,博士生,主要从事三维重建、视觉位姿测量、目标识别等方面的研究。通信地址:哈尔滨工业大学深空探测基础研究中心(150080),电话:(0451)86402628,E-mail:yuanmengmeng_HIT@163.com;通信作者:张泽旭(1971-),男,博士生导师,教授,主要从事飞行器自主导航控制、无人机智能协同感知决策以及数据可视化等方面的研究。通信地址:哈尔滨工业大学深空探测基础研究中心(150080),电话:(0451)86402628,E-mail:zexuzhang@hit.edu.cn。
