FPN算法在视觉感知机器人抓取控制的应用研究  

Research on the Application of FPN Algorithm in Grasping Control of Visual Perception Robot

作　　者：王利祥[1] 郭向伟 卢明星[1] WANG Li-xiang;GUO Xiang-wei;LU Ming-xing(Public Discipline Department of He’nan Vocational College of Nursing,He’nan Anyang 455000,China;School of Electrical Engineering and Automation,He’nan University of Technology,He’nan Jiaozuo 454000,China)

机构地区：[1]河南护理职业学院公共学科部,河南安阳455000 [2]河南理工大学电气工程与自动化学院,河南焦作454000

出　　处：《机械设计与制造》2024年第4期303-307,313,共6页Machinery Design & Manufacture

基　　金：2022年河南省医学教育研究项目(Wjlx2022272)—基于有效教学理论的混合式课堂教学模式研究与实践—以《医用计算机基础》。

摘　　要：针对视觉感知机器人对物体抓取的准确性控制,在抓取姿势估计基础上使用密集连接的特征金字塔网络(FPN)作为特征提取器,将语义更强的高级特征图与分辨率更高的低级特征图融合,将机器人物体抓取过程分为两个阶段,第一个阶段生成待抓取区域,第二阶段对抓取区域进行细化以预测抓取姿势。模型在Cornell抓取数据集和Jacquard数据集上训练,验证了所提算法在抓取姿势估计的有效性。设计了两种不同真实场景的物体抓取控制实验,结果表明所提模型能有效提高机器人抓取各种不同尺寸物体的能力。Aiming at the accuracy control of object grasping by visual perception robot,on the basis of grasping pose estimation,the densely connected feature pyramid network(FPN)is used as the feature extractor to fuse the high-level feature map with stron-ger semantics and the low-level feature map with higher resolution.The grasping process of the robot human body is divided into two stages:the first stage generates the area to be grasped,and the second stage refines the grasping area to predict the grasping pose.The model is trained on Cornell and Jacquard data sets,which verifies the effectiveness of the proposed algorithm in grasp-ing pose estimation.Two kinds of real scene object grasping control experiments are designed.The results show that the proposed model can effectively improve the robot's ability to grasp objects of different sizes.

关 键 词：视觉机器人 抓取姿势 FPN 特征图融合 

分 类 号：TH16[机械工程—机械制造及自动化] TP242.2[自动化与计算机技术—检测技术与自动化装置]