一种基于轮廓特征匹配的零件位姿估计方法  

Research of Parts Pose Estimation Method Based on Boundary Feature Matching

作　　者：孙长江[1,2] 段旭洋[1,2] 王皓[1,2,3] 陈智超[4] 徐鹏 SUN Changjianngg;DUAN Xuyang;WANG Haoo;CHEN Zhichao;XU Peng(Shanghai Key Laboratory of Digital Manufacture for Thin-walled Structures,Shanghai Jiao Tong University,Shanghai 200240,China;Fraunhofer Innovation Platform for Smart Manufacturing at Shanghai Jiao Tong University,Shanghai 201306,China;Key Laboratory of Mechanical System and Vibration,Shanghai Jiao Tong University,Shanghai 200240,China;5G Industry Innovation Center,Shanghai Aircraft Manufacturing Co.,Ltd,Shanghai 201202,China)

机构地区：[1]上海交通大学上海市复杂薄板结构数字化制造重点实验室,上海200240 [2]上海交通大学弗劳恩霍夫协会智能制造创新中心,上海201306 [3]上海交通大学机械系统与振动国家重点实验室,上海200240 [4]上海飞机制造有限公司5G工业创新中心,上海201202

出　　处：《机械设计与研究》2023年第3期109-114,共6页Machine Design And Research

基　　金：临港地区智能制造产业专项(ZN2017020102)。

摘　　要：针对现有位姿估计算法在工业零件散乱堆叠场景下精度差、效率低的问题,提出了一种采用轮廓特征匹配的三维目标位姿估计方法。利用场景点云数据生成场景深度图像,采用Sobel滤波器提取图像边缘并得到场景点云轮廓;提出基于主方向一致性的轮廓点云聚类方法,对不同边缘的轮廓点云聚类并提取直线轮廓特征,建立用于进行轮廓特征匹配的统一特征描述子;通过哈希表存储的方式建立模型特征描述表,降低特征匹配时间,并通过霍夫投票得到多个候选位姿;使用位姿聚类的方式对候选位姿进行评价和筛选从而得到最终的位姿结果。采用工业目标识别数据集ITODD和自采数据进行目标6D位姿估计实验,实验结果表明提出的位姿估计方法相比于原始点对特征匹配算法识别速度更快,准确率更高,算法单个场景平均识别速度3.78 s,平均识别准确率76%。Aiming at the problems of poor accuracy and low efficiency of existing pose estimation algorithms in the scene of scattered industrial parts,this paper proposes a pose estimation method using boundary features.Firstly,the scene point cloud is used to generate a depth image,and the Sobel filter is used to extract the edges of the image and obtain the scene point cloud boundary.A point cloud clustering method based on the consistency of the main direction is proposed to cluster the boundary point clouds of different edges and extract the line features.A unified feature descriptor is established for boundary feature matching;a model feature description table is established to reduce feature matching time,and multiple candidate poses are obtained through Hough voting.The candidate poses are evaluated and screened through pose clustering to obtain the final pose results.The industrial target recognition data set ITODD and self-collected data are used to perform the target 6D pose estimation experiment.The experimental results show that the proposed pose estimation method has faster recognition speed and higher accuracy than the original point pair feature matching algorithm.The average recognition speed of a single scene point cloud is 3.78 s,and the average recognition accuracy is 76%.

关 键 词：位姿估计 点云轮廓提取 特征匹配 工业零件分拣 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置]