Abstract: Aiming at the problem that the lack of regular features of large free-form parts leaded to high positioning cost,low efficiency and difficult to guarantee accuracy,a rapid robot vision positioning method was proposed.An adaptive resolution supervoxel clustering algorithm was designed to uniformly segment point clouds of different models.Based on the curvature,angle and Fast Point Feature Histogram (FPFH) descriptor of the block centroid points,the overlap area between the complete CAD model point cloud and the local measurement data was identified quickly and the coarse registration matrix was solved.The Branch and Bound (BNB) nested Iterative Closest Point (ICP) algorithm was improved,and a fast fine registration method based on key points was presented.Experimental verification was performed on stamping parts,and the results showed that the proposed method saved more than 30% of the registration time,and significantly improved the registration accuracy,which could solve the positioning problem of large free-form surface parts.

Key words: free-form surface parts, robot vision positioning, division, centroid point, overlapping region, key point

摘要： 针对大型自由曲面零件缺少规则特征导致定位成本高、效率低且精度难以保证的问题,提出一种机器人视觉快速定位方法。设计了自适应分辨率的超体素聚类算法,对不同模型点云进行均匀分割;提出以分块质心点的曲率、夹角以及快速点特征直方图(FPFH)描述子为依据,快速识别出完整CAD模型点云与局部测量数据的重叠区域并求解粗配准矩阵。改进了分支定界(BNB)嵌套迭代最近点(ICP)算法,提出基于关键点的快速精配准方法。以冲压件为对象进行实验验证,结果表明本方法节省了30%以上的配准时间,并显著提高了配准精度,能够解决大型自由曲面零件的定位难题。

关键词: 自由曲面零件, 机器人视觉定位, 分块, 质心点, 重叠区域, 关键点