首页 | 本学科首页   官方微博 | 高级检索  
     

基于三维检测网络的机器人抓取方法
引用本文:葛俊彦,史金龙,周志强,王 直,钱 强.基于三维检测网络的机器人抓取方法[J].仪器仪表学报,2021(8):146-153.
作者姓名:葛俊彦  史金龙  周志强  王 直  钱 强
作者单位:1.江苏科技大学计算机学院
基金项目:科技部国家重点研发计划(2018YFC0309104)项目资助
摘    要:机器人抓取任务中面对的是不同形状和大小的物体,而散落在场景中的物体会有不同的姿态和位置,这对机器人抓取中计算物体位姿任务提出了较高的挑战。针对于此,本文设计了一种基于三维目标检测的机器人抓取方法,弥补了基于二维图像识别引导机器人抓取任务中对视角要求较高的缺陷。首先,设计了一种卷积神经网络在RGB图像中识别物体,并回归出物体三维包围盒、物体中心点;其次,提出一种计算机器人抓取物体最佳姿势的策略;最后,控制机器人进行抓取。在实际场景中,使用本文设计的三维检测网络,三维目标检测精度达到88%,抓取成功率达到94%。综上所述,本文设计的系统能有效找到机器人合适的抓取姿势,提高抓取成功率,满足更高的抓取任务要求。

关 键 词:深度学习  三维检测  位姿计算  机器人控制

A robotic grasping method based on three-dimensional detection network
Ge Junyan,Shi Jinlong,Zhou Zhiqiang,Wang Zhi,Qian Qiang.A robotic grasping method based on three-dimensional detection network[J].Chinese Journal of Scientific Instrument,2021(8):146-153.
Authors:Ge Junyan  Shi Jinlong  Zhou Zhiqiang  Wang Zhi  Qian Qiang
Affiliation:1.School of Computer Science and Engineering, Jiangsu University of Science and Technology;2.School of Computer Science and Engineering, Jiangsu University of Science and Technology;3.School of Computer Science and Engineering, Jiangsu University of Science and Technology;4.School of Computer Science and Engineering, Jiangsu University of Science and Technology; 5.School of Computer Science and Engineering, Jiangsu University of Science and Technology
Abstract:The robot faces different shapes and sizes of objects in the task of grasping. The scattered objects in the scene may have different poses and positions, which make the task of recognizing positions and poses of objects more difficult. In view of this, a threedimensional scene recognition method for robotic grasping is proposed. It makes up a defect that the 2D detection method is sensitive to the field of view in robotic grasping task. Firstly, the convolutional neural network is designed to detect the object in the RGB image. Eight corner points of the three-dimensional bounding box of objects, and the center point of the object are generated. Secondly, a method is proposed to calculate the best position and pose for robotic grasping. Finally, the robot is controlled to grasp objects. In real scene, the detection accuracy reaches 88% , and the grasping success rate based on the designed three-dimensional recognition network is up to 94% . In summary, the designed network can effectively find a suitable grasping pose. The grasping success rate is improved. It is able to meet higher requirements.
Keywords:deep learning  3D detection  pose calculation  robot control
本文献已被 CNKI 等数据库收录!
点击此处可从《仪器仪表学报》浏览原始摘要信息
点击此处可从《仪器仪表学报》下载全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号