摘要
针对机器人无序抓取场景中,因工件堆叠遮挡和对称结构相似等因素导致的位姿估计不精准等问题,提出了一种基于加权余弦相似度投票的点对特征位姿估计算法。在离线阶段,采用虚拟多视角提取点云构建全局模型,提高后续匹配效率;在线匹配阶段,提出一种多分辨率体素网格下采样方法,精简点云数量并保留关键点。经过广义霍夫投票生成候选位姿后,提出欧式距离结合余弦相似度的加权投票方法,以筛选出最优姿势,实现精确的目标位姿估计。在公共数据集和自制点云数据的测试结果表明,与点对特征算法和快速点对特征算法相比,该算法在平均召回率上提升了20%和14%,识别率和运行效率均有提升。
In order to solve the problem of imprecision pose estimation caused by workpiece stack occlusion and symmetrical structure similarity,a weighted cosine similarity voting algorithm for point pair feature pose estimation was proposed.In the offline stage,the virtual multi-view point cloud is extracted to build the global model and improve the subsequent matching efficiency.In the online matching stage,a multi-resolution voxel grid subsampling method is proposed to simplify the number of point clouds and retain the key points.After the candidate pose is generated by generalized Hough voting,the weighted voting method of Euclidean distance combined with cosine similarity is proposed to select the best pose and achieve accurate target pose estimation.The test results on public data sets and self-made point cloud data show that compared with point-to-feature algorithm and fast point-to-feature algorithm,the average recall rate of this algorithm is improved by 20%and 14%,and the recognition rate and operation efficiency are significantly improved.
作者
王一
崔振浩
程佳
付智超
WANG Yi;CUI Zhenhao;CHENG Jia;FU Zhichao(College of Electrical Engineering,North China University of Science and Technology,Tangshan 063210,China;Key Laboratory of Advanced Detection and Control Technology in Tangshan City,North China University of Science and Technology,Tangshan 063210,China)
出处
《组合机床与自动化加工技术》
北大核心
2025年第7期89-94,共6页
Modular Machine Tool & Automatic Manufacturing Technique
基金
河北省高等学校科学技术研究项目(ZD2022114)
唐山市应用基础研究项目(21130212C)
教育部产学合作协同育人项目(220804992272302)。
关键词
位姿估计
点对特征
多分辨率体素
余弦相似度
加权投票
pose estimation
point pair feature
multi-resolution voxel
cosine similarity
weighted voting
作者简介
王一(1981-),男,教授,博士,研究方向为机器视觉感知和工业机器人应用技术,(E-mail)wangyi@ncst.edu.cn;通信作者:崔振浩(1999-),男,硕士研究生,研究方向为机器视觉和三维重建,(E-mail)cuizhenghao0215@163.com。
