Abstract: The positioning and grabbing of robotic targets are central to the field of industrial robotic automation. In order to solve the problem of time-consuming and error-prone 3D robot grasping, this paper proposes a machine vision grasping scheme based on CICP registration and establishes a 3D multi-dimensional space target model location computer vision system. The rough area of the gradient histogram feature+support vector machine is used to get the approximate target area, then the target contour is extracted by using the preprocessing method based on Guide filtering and flood fill to get the contour point cloud. ICP is used for point cloud registration to obtain the perspective transformation matrix of the target workpiece, which can well solve the problem of locating and grasping the multi-target workpiece on the conveyor belt. The experiment shows that the system can control the grasping error within 0.4 mm, and for small workpieces, the registration time can be shortened to less than 200 ms.

Key words: multi-target positioning and grabbing, histogram of oriented gradient, contour iterative closest point

摘要： 机械臂的目标定位和抓取是工业机器人自动化领域的核心问题。针对目前机器人3D抓取耗时长且误差较大的问题，提出基于轮廓迭代最近点（Contour Iterative Closest Point，CICP）配准的机器视觉抓取方案，并建立多维空间下的3D多目标模型定位的计算机视觉系统。利用梯度直方图特征+支持向量机对多目标进行检测分割得到单目标大致区域，再使用基于Guide滤波+漫水填充的预处理方式对目标轮廓进行提取得到轮廓点云，最后使用迭代最近点（Iterative Closest Point，ICP）进行点云配准获取目标工件的透视变换矩阵，能够很好地解决传送带上多目标工件的定位与抓取问题，实验结果证明该系统抓取误差可降低到0.4 mm，且对于小工件可将配准时间缩短至200 ms以内。

关键词: 多目标定位与抓取, 梯度直方图特征, 轮廓迭代最近点配准