摘要
针对局部自主遥操作过程中识别目标准确率低的问题,提出了一种基于改进快速区域卷积神经网络的抓取构型识别方法,通过对其区域生成网络中锚点尺度、前景特征区域、候选框的线性回归和分类网络分别进行改进,以提高抓取构型识别的准确率。首先将抓取构型参数化,然后在目标区域中利用锚点法对抓取构型参数进行识别,结合视觉传感器采集到的深度信息确定目标高度,并通过线性回归方法对抓取区域进行修正。通过搭建机器人试验平台,利用Cornell Grasp Dataset进行训练与测试进行验证。试验结果表明,提出的方法在简单网络识别准确率为96.4%,并成功实现机器人对目标的抓取。
Based on improved faster regional convolutional neural network(faster R-CNN),an object detection and grasp configuration reorganization method was proposed to solve the problem of low accuracy in local autonomy teleoperation.The anchor scale of the regional proposal network,the foreground feature area,the candidate box linear regression and the classification network were improved to increase the accuracy of the grasp configuration recognition.First,the grasp configuration was parameterized by the algorithm.Then the anchor method was used to identify the grasp configuration parameters in the target area,and the depth information collected by the vision was incorporated to determine the object height.Through the linear regression method,the grasping area was corrected to make the grasp configuration parameter identification more precise,and the success rate of the capture was improved.In addition,a robot experiment platform was built and the network was trained and tested using the Cornell Grasp Dataset data set.The results showed that the proposed method achieved 96.4%accuracy in a simple network,and successful grasp of target by the robot was achieved.
作者
韩冬
黄攀峰
齐志刚
HAN Dong;HUANG Panfeng;QI Zhigang(Research Center of Intelligent Robotics,Northwestern Polytechnical University,Xi’an 710072,China;National Key Laboratory of Aerospace Flight Dynamics,Northwestern Polytechnical University,Xi’an 710072,China;School of Physics and Information Engineering,Shanxi Normal University,Linfen 041000,China)
出处
《载人航天》
CSCD
北大核心
2019年第5期586-593,共8页
Manned Spaceflight
基金
载人航天预先研究项目(2018KC020081)
