期刊检索

  • 2019年第51卷
  • 2018年第50卷
  • 2017年第49卷
  • 2016年第48卷
  • 2015年第47卷
  • 2014年第46卷
  • 2013年第45卷
  • 2012年第44卷
  • 2011年第43卷
  • 2010年第42卷
  • 第1期
  • 第2期

主管单位 中华人民共和国
工业和信息化部
主办单位 哈尔滨工业大学 主编 冷劲松 国际刊号ISSN 0367-6234 国内刊号CN 23-1235/T

期刊网站二维码
微信公众号二维码
引用本文:任秉银,魏坤,代勇.一种非结构环境下目标识别和3D位姿估计方法[J].哈尔滨工业大学学报,2019,51(1):38-44.DOI:10.11918/j.issn.0367-6234.201804210
REN Bingyin,WEI Kun,DAI Yong.A novel method of target recognition and 3D pose estimation in unstructured environment[J].Journal of Harbin Institute of Technology,2019,51(1):38-44.DOI:10.11918/j.issn.0367-6234.201804210
【打印本页】   【HTML】   【下载PDF全文】   查看/发表评论  下载PDF阅读器  关闭
过刊浏览    高级检索
本文已被:浏览 134次   下载 191 本文二维码信息
码上扫一扫!
分享到: 微信 更多
一种非结构环境下目标识别和3D位姿估计方法
任秉银,魏坤,代勇
(哈尔滨工业大学 机电工程学院, 哈尔滨 150001)
摘要:
为提高非结构环境下目标识别准确率和位姿估计精度,提出一种利用Kinect V2 RGB-D传感器并基于目标的CAD模型进行不同类型目标自动识别和3D位姿估计的新方法.利用虚拟相机获取目标CAD模型的深度图像,并将目标的模型转化为点云图,采用体素栅格滤波减少场景点云中的点数;利用点对特征描述子(PPF)作为CAD模型的全局描述子,并将相似的PPF划分成一组放进一个hash表,用于识别和定位目标,所有目标的hash表组成了3D模型数据库;利用基于投票策略的方法对不同类型目标进行检测识别和3D位姿估计,并采用位姿聚类的方法和ICP配准进行位姿修正,再通过奇异值滤波剔除误匹配位姿,从而提高位姿估计精度. 在虚拟机器人实验平台仿真环境中分析了3种管接头的识别率和位姿估计误差,结果表明:3种管接头平均识别率96%,位置误差<4 mm,姿态误差<2°,能够满足机械臂抓取要求. 将提出的方法与两种主流位姿估计方法进行了对比实验,结果表明,提出的方法无论是识别率还是F1分数都要优于其他两种方法.
关键词:  目标识别  3D位姿估计  非结构环境  CAD模型  位姿聚类  ICP配准
DOI:10.11918/j.issn.0367-6234.201804210
分类号:TP391
文献标识码:A
基金项目:
A novel method of target recognition and 3D pose estimation in unstructured environment
REN Bingyin,WEI Kun,DAI Yong
(School of Mechatronics Engineering, Harbin Institute of Technology, Harbin 150001, China)
Abstract:
To improve target recognition accuracy and pose estimation precision in unstructured environment, a novel method for automatic recognition and 3D pose estimation of different kinds of targets is presented based on the object CAD model and Kinect V2 RGB-D sensor. The depth image of object CAD model is obtained by a virtual camera, and it is converted into one point cloud. A voxel grid filter is utilized to reduce the number of points for the point cloud of the scene, and the point pair feature(PPF) is used as the global descriptor of CAD model. The similar PPFs are classified as the same group and put into one hash table to recognize and locate the targets. The hash tables of all targets compose the 3D model database. A voting scheme is adopted for different kinds of objects recognition and 3D pose estimation. Pose cluster and ICP registration are employed to refine 3D pose. The accuracy of 3D pose estimation is improved by filtering the mismatching poses. The recognition rate and the pose estimation error of three kinds of pipe joints are analyzed in the environment of virtual robot experimentation platform (V-REP). The simulation results show that the average recognition rate of three kinds of pipe joints is 96%. The position error is less than 4 mm and the orientation error is no more than 2°, which meet the requirement of manipulator grasping. The experiments of comparison with other mainstream methods are performed. The experimental results show that the recognition rate and the F1 score of the proposed method are superior to those of two other methods.
Key words:  target recognition  3D pose estimation  unstructured environment  CAD model  pose clustering  ICP registration

友情链接LINKS