基于实例分割的机械臂抓取位姿检测方法  

Robotic Arm Gripping Position Detection Method Based on Instance Segmentation

作　　者：翟维枫 陆文涛 薛同来 ZHAI Weifeng;LU Wentao;XUE Tonglai(College of Electrical and Control Engineering,North China University of Technology)

机构地区：[1]北方工业大学电气与控制工程学院

出　　处：《仪表技术与传感器》2024年第10期89-94,116,共7页Instrument Technique and Sensor

基　　金：国家自然科学基金面上项目(62173002);北京市自然科学基金面上项目(4222045)。

摘　　要：为了实现机械臂对常见物体的高效抓取,提出一种将实例分割网络与抓取位姿检测(GPD)算法结合,并加入抓取位姿筛选的抓取位姿检测方法。通过YOLOv8s-Seg算法在RGB图像中分割出抓取目标物体,结合深度图像生成物体点云,使用GPD法检测点云得到候选抓取位姿,对候选抓取位姿进行评分,取分数最高的抓取位姿规划执行抓取。建立了基于机器人操作系统(ROS)的3D视觉机械臂抓取系统,并运用所提方法进行物体抓取实验。实验结果表明所提方法的抓取成功率较GPD法有显著提升。In order to realize efficient grasping of common objects by robotic arms,a grasping position detection method that combines an instance segmentation network with grasp position detection(GPD)algorithm and incorporates grasping position filtering was proposed.The YOLOv8s-Seg algorithm segmented the grasping target object in the RGB image,combined it with the depth image to generate the object point cloud,detected the point cloud using the GPD algorithm to get the candidate grasping poses,scored the candidate grasping poses,and took the grasping poses with the highest scores to plan the execution of the grasping.A 3D vision robotic arm grasping system based on robot operating system(ROS)was established and the proposed method was deployed for object grasping experiments,and the experimental results show that the grasping success rate of the proposed method is significantly improved compared with the GPD method.

关 键 词：实例分割 点云 GPD 抓取位姿检测 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]