基于3D视觉的机械零件位姿分析  

Pose Analysis of Mechanical Parts Based on 3D Vision

作　　者：王洪申[1] 曹熙淦 李昌德 WANG Hongshen;CAO Xigan;LI Changde(School of Mechanical and Electrical Engineering,Lanzhou University of Technology,Lanzhou 730050,China)

机构地区：[1]兰州理工大学机电工程学院,兰州730050

出　　处：《组合机床与自动化加工技术》2023年第8期131-134,共4页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：国家自然科学基金项目(61962035)。

摘　　要：针对工业制造领域中存在的弱纹理或无纹理的机械零件位姿估计问题,以DenseFusion网络为基础,提出了一种基于3D视觉的机械零件位姿估计方法。首先,依照工业制造场景构建了用于网络训练和测试的机械零件仿真数据集;其次,使用目标零件分割出的深度图像构建点云,对其进行曲率下采样处理,提高关键点的质量;最后,将颜色特征、点云特征和法线特征融合,使用融合特征回归目标零件的6D位姿。在构建的机械零件仿真数据集和LineMod公共数据集进行了实验和比较,其结果表明了所提出的方法相较于同类其他方法具有更高的准确率,更好的收敛性能,对于弱纹理或无纹理的机械零件有较好的位姿估计效果。Aiming at the pose estimation problem of weak texture or no texture of mechanical parts in the field of industrial manufacturing,this paper proposes a pose estimation method for mechanical parts based on 3D vision based on DenseFusion network.Firstly,a mechanical parts simulation dataset for network training and testing is constructed according to the industrial manufacturing scenario.Secondly,the point cloud is constructed using the depth image segmented by the target part,curvature downsampling are used to improve the quality of key points.Finally,the color features,point cloud features,and normal features are fused,and the fusion features are used to return to the 6D pose of the target part.The results show that the proposed method has higher accuracy,better conver-gence performance,and better pose estimation effect for weakly textured or untextured mechanical parts than other similar methods.

关 键 词：位姿估计 深度学习 3D视觉 特征提取 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]