摘要
为了提高机器人在抓取物体时成功率和速度,提出了一种基于SE-ResNet的生成式机器人抓取位置检测模型。模型通过输入的RGB-D图像直接生成对应的抓取映射图,ResNet解决了深度模型的退化问题,在此基础上引入挤压和激励机制,让模型使用全局信息来增强有用信息,并抑制无用信息。选择Jaccard指数作为抓取检测的评价标准。实验结果表明,该方法在Cornell数据集上的准确率高达98.9%。搭建了基于PyBullet物理仿真环境的抓取检测平台,使用Panda机械臂对Egad评估集中的各个难度的物体抓取表现良好,在提高抓取检测的成功率的同时,保证了实时性。
In order to improve the success rate and speed of robot grasping,a generative robot grasping position detection model based on SE-ResNet is proposed.The model generates the corresponding grasping map directly from the input RGB-D image.ResNet solves the degradation problem of depth model.On this basis,it introduces squeeze and excitation mechanism to make the model use global information to enhance useful information while suppress useless information.Jaccard index is selected as the evaluation standard of grasping detection.The experimental results show that the accuracy of this method is up to 98.9%on the Cornell dataset.A physical simulation platform of grasping detection based on PyBullet is built,and the Panda manipulator is used to grasp the objects with different difficulty in Egad evaluation set,which assures real-time performance at the same time increases success rate of grasping detection.
作者
杨华
宋卓著
吴杰宏
高利军
YANG Hua;SONG Zhuozhu;WU Jiehong;GAO Lijun(School of Computer Science,Shenyang Aerospace University,Shenyang 110136,China)
出处
《传感器与微系统》
CSCD
北大核心
2023年第5期112-116,共5页
Transducer and Microsystem Technologies
基金
辽宁省自然科学基金资助项目(2019—ZD—0243)
辽宁省教育厅科学技术研究项目(L201626)
航空科学基金资助项目(2018ZC54013)。
关键词
抓取检测
机器人抓取
深度学习
挤压与激励
grasping detection
robotic grasp
deep learning
squeeze and excitation