基于三流联合卷积神经网络的机械臂抓取检测  

Robotic Grasping Detection Based on Joint Convolution Neural Network of Three Streams

作　　者：王勇[1] 陈荟西 WANG Yong;CHEN Hui-xi(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《小型微型计算机系统》2020年第5期1112-1116,共5页Journal of Chinese Computer Systems

基　　金：国家自然科学青年基金项目(61502065)资助;重庆市科学技术委员会基础与前沿研究重点项目(cstc2015jcyjBX0127)资助;重庆市巴南区技术合作项目[2016]33资助.

摘　　要：目前已有一些研究将深度学习应用于机械臂的抓取检测中,但很难同时保证检测的实时性和准确性.本文提出一种用于抓取检测的三流联合卷积神经网络模型,该模型由三个独立的深度卷积神经网络流组成,其中每个流分别处理其对应的输入模态,并将它们以一种后期融合的方式结合在一起;然后利用改进的单级回归算法进行抓取位置预测,并提出一种新的置信度计算方式.该模型在康奈尔抓取数据集上图像分割和对象分割的准确率分别为94.9%和93.7%.并且在GPU上以每秒14.2帧的速度进行实时检测.测试结果表明,我们的模型同时保证了抓取检测的实时性与准确性,提高了检测的速度与精度.At present,some researches has been applied deep learning to robotic grasping detection,but it is difficult to guarantee its real-time speeds and accuracy at the same time.In this paper,a joint convolution neural network(CNN)of three streams model is proposed for grasping detection.The model is composed of three independent streams with deep CNNs,each stream processes its corresponding input mode separately,and combined these in a late fusion approach;Then use improved single-stage regression algorithm for grasping position prediction.The accuracy of image-wise split and object-wise split on the Cornell Grasp Detection Dataset is 94.9% and 93.7% respectively,and the model runs at 14.2 frames per second on a GPU for real-time detection.The test results show that our model can guarantee the real-time speeds and accuracy of grasping detection,and improve the efficiency and performance of grasping.

关 键 词：抓取检测 三流联合 深度卷积神经网络 单级回归 置信度计算 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]