基于双视角点云拼接的机械手抓取方法  

Grasping method of manipulator based on two registered point clouds

作　　者：张自杰 张国良 曾静 李歆 谢波 Zhang Zijie;Zhang Guoliang;Zeng Jing;Li Xin;Xie Bo(Artificial Intelligence Key Laboratory of Sichuan Province,Yibin 644000,China;School of Automation and Information Engineering,Sichuan University of Science and Engineering,Yibin 644000,China)

机构地区：[1]人工智能四川省重点实验室,宜宾644000 [2]四川轻化工大学自动化与信息工程学院,宜宾644000

出　　处：《国外电子测量技术》2022年第11期102-108,共7页Foreign Electronic Measurement Technology

基　　金：四川省应用基础研究项目(2019YJ00413)资助。

摘　　要：针对传统的抓取方法难以获得可靠抓取位姿的问题,提出一种基于双视角点云拼接的抓取方法。首先,通过手眼标定将两个视角下的点云同时转化到机器人基坐标系下,实现点云粗拼接;其次,为了解决迭代最近点(ICP)算法容易陷入局部最优的问题,设计了一种基于垂直平面的ICP算法(VP-ICP),引入辅助参照物以补偿错误匹配点所带来的计算误差,实现精细拼接;然后,对抓取位姿检测(GPD)算法进行改进,调整候选抓取位姿的筛选条件,提高对贴近地面的物体的抓取能力;最后,搭建了一套基于机器人操作系统(ROS)的3D视觉抓取系统,使用改进的GPD法在拼接后的点云中检测出可靠的抓取位姿,实现对散乱堆放物体的无序抓取。实验结果表明,所提算法的平均拼接耗时为1.23 s,地面法向量夹角的均值为0.56°,较次优算法有明显提升;点云拼接和改进的GPD法分别将平均抓取完成度提升了12.5%和4.1%。To solve the problem that it is difficult to obtain reliable grasp poses by traditional grasping methods, a new grasping algorithm based on two registered point clouds is presented in this paper. Firstly, two point clouds captured from different observation perspectives are both transformed to base frame for coarse registration, which is based on hand-eye calibration. Secondly, to tackle the problem that ICP algorithm is easy to fall into local optimal solution, an improved vertical plane-based ICP method named VP-ICP is designed to compensate the calculation errors caused by wrong paired points. VP-ICP method is implemented by introducing assistance reference objects and used for fine registration. Then, GPD algorithm is improved by modifying the conditions of refining grasp candidates to enhance its capacity for grasping objects close to ground. Finally, a 3 D vision-based grasping system implemented by ROS is constructed to accomplish bin-picking for cluttered objects using the improved GPD methods which can find reliable grasp poses in the registered point cloud. The experiment results show that the presented method gets average registration time of 1.23 s and 0.56° average angle between ground normal, significantly better than the second-best method, and average grasp completion rate is respectively increased by 12.5% and 4.1% through registration and improved GPD.

关 键 词：点云拼接 手眼标定 迭代最近点算法 GPD 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置]