基于语义分割与旋转目标检测的机器人抓取位姿估计  

Robot grasping pose estimation based on semantic segmentation androtating target detection

作　　者：孙先涛 闻勇 陈文杰 智亚丽 胡祥涛 陈伟海[2] SUN Xian-tao;WEN Yong;CHEN Wen-jiey;ZHI Ya-li;HU Xiang-tao;CHEN Wei-hai(School of Electrical Engineering and Automation,Anhui University,Hefei 230601,China;School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]安徽大学电气工程与自动化学院,合肥230601 [2]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《控制与决策》2024年第9期2913-2922,共10页Control and Decision

基　　金：国家自然科学基金项目(52005001)。

摘　　要：针对姿态任意、尺寸不一的物体以及抓取角度离散性问题,提出一种基于语义分割与旋转目标检测的单目位姿估计方法.阶段1:首先利用faster R-CNN (faster regions with convolutional neural networks features)进行抓取检测获取候选抓取框;其次利用语义分割网络获取待抓取物体的轮廓信息;最后利用语义分割结果为每个待抓取物体筛选置信度最高的抓取框,同时完成角度粗估计.阶段2:利用旋转目标检测获取精细的抓取角度,以修正阶段1抓取框的偏转角.此外,考虑到抓取物具有多尺度的特点,提出一种多尺度特征融合模块,使金字塔的所有层共享相似的语义特征.针对智能算法求解逆运动学时出现迭代后期收敛速度慢的问题,利用牛顿法收敛速度快的优点,加快智能算法后期的收敛速度.基于V-REP仿真以及实际抓取检测实验表明,所提算法的抓取检测精度为98.4%,实际抓取成功率达到了88.3%,仿真抓取时的有害扭矩大小较修正前有所改善,能够满足机械臂抓取要求.A monocular pose estimation method based on semantic segmentation and rotating target detection is proposed to solve the problem of arbitrary pose,different size objects and discrete grasping angles.In the first stage,faster R-CNN(faster regions with convolutional neural networks features)is used for grasping detection to obtain candidate grasping boxes.Then,the semantic segmentation network is used to obtain the contour information of the object to be captured.Finally,the semantic segmentation results are used to select the most reliable grasping box for each object to be grasped,and the coarse angle estimation is completed at the same time.In the second stage,the fine grasping angle is obtained by rotating target detection to correct the deflection angle of the grasping frame in the first stage.In addition,considering the multi-scale characteristics of the grab,a multi-scale feature fusion module is proposed to make all layers of the pyramid share similar semantic features.Aiming at the problem of slow convergence speed in the later stage of iteration when the intelligent algorithm solves the inverse kinematics,the advantage of fast convergence speed of the Newton method is used to speed up the convergence speed of the intelligent algorithm in the later stage.Based on V-REP simulation and actual grasping detection experiments,the grasping detection accuracy of the proposed algorithm is 98.4%,and the actual grasping success rate is 88.3%.The harmful torque during simulation grasping is improved compared with that before correction,which can meet the grasping requirements of the manipulator.

关 键 词：抓取检测 语义分割 特征金字塔 V-REP仿真 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]