现在位置: > > IT/计算机 > 计算机软件及应用

Real-Time Grasp Detection Using Convolutional Neural Networks

Real-Time Grasp Detection Using Convolutional Neural Networks

Real-TimeGraspDetectionUsingConvolutionalNeuralNetworks

JosephRedmon1,AneliaAngelova2

Abstract—Wepresentanaccurate,real-timeapproachtoroboticgraspdetectionbasedonconvolutionalneuralnetworks.Ournetworkperformssingle-stageregressiontograspableboundingboxeswithoutusingstandardslidingwindoworregionproposaltechniques.Themodeloutperformsstate-of-the-artapproachesby14percentagepointsandrunsat13framespersecondonaGPU.Ournetworkcansimultaneouslyperformclassi cationsothatinasinglestepitrecognizestheobjectand ndsagoodgrasprectangle.Amodi cationtothismodelpredictsmultiplegraspsperobjectbyusingalocallyconstrainedpredictionmechanism.Thelocallyconstrainedmodelperformssigni cantlybetter,especiallyonobjectsthatcanbegraspedinavarietyofways.

I.INTRODUCTION

Perception—usingthesenses(orsensorsifyouarearobot)tounderstandyourenvironment—ishard.Visualper-ceptioninvolvesmappingpixelvaluesandlightinformationontoamodeloftheuniversetoinferyoursurroundings.Gen-eralsceneunderstandingrequirescomplexvisualtaskssuchassegmentingasceneintocomponentparts,recognizingwhatthosepartsare,anddisambiguatingbetweenvisuallysimilarobjects.Duetothesecomplexities,visualperceptionisalargebottleneckinrealroboticsystems.

Generalpurposerobotsneedtheabilitytointeractwithandmanipulateobjectsinthephysicalworld.Humansseenovelobjectsandknowimmediately,almostinstinctively,howtheywouldgrabthemtopickthemup.Roboticgraspdetectionlagsfarbehindhumanperformance.Wefocusontheproblemof ndingagoodgraspgivenanRGB-Dviewoftheobject.

WeevaluateontheCornellGraspDetectionDataset,anextensivedatasetwith

Real-Time Grasp Detection Using Convolutional Neural Networks

numerousobjectsandground-truthlabelledgrasps(seeFigure1).Recentworkonthisdatasetrunsat13.5secondsperframewithanaccuracyof75percent[1][2].Thistranslatestoa13.5seconddelaybetweenarobotviewingasceneand ndingwheretomoveitsgrasper.Themostcommonapproachtograspdetectionisaslidingwindowdetectionframework.Theslidingwindowapproachusesaclassi ertodeterminewhethersmallpatchesofanimageconstitutegoodgraspsforanobjectinthatimage.Thistypeofsystemrequiresapplyingtheclassi ertonumerousplacesontheimage.Patchesthatscorehighlyareconsideredgoodpotentialgrasps.

Wetakeadifferentapproach;weapplyasinglenetworkoncetoanimageandpredictgraspcoordinatesdirectly.Ournetworkiscomparativelylargebutbecauseweonlyapplyitoncetoanimagewegetamassiveperformanceboost.

1UniversityofWashington2Google

Research

Fig.1.TheCornellGraspingDatasetcontainsavarietyofobjects,eachwithmultiplelabelledgrasps.Graspsaregivenasorientedrectanglesin2-D.

Insteadoflookingonlyatlocalpatchesournetworkusesglobalinformationintheimagetoinformitsgrasppredic-tions,makingitsigni cantlymoreaccurate.Ournetworkachieves88percentaccuracyandrunsatreal-timespeeds(13framespersecond).Thisrede nesthestate-of-the-artforRGB-Dgraspdetection.

II.RELATEDWORK

Signi cantpastworkuses3-Dsimulationsto ndgoodgrasps[3][4][5][6][7].Theseapproachesarepowerfulbutrelyonafull3-Dmodelandotherphysicalinformationaboutanobjectto ndanappropriategrasp.Fullobjectmodelsareoftennotknownapriori.Generalpurposerobotsmayneedtograspnovelobjectswithout rstbuildingcomplex3-Dmodelsoftheobject.

RoboticsystemsincreasinglyleverageRGB-Dsensorsanddatafortaskslikeobjectrecognition[8],detection[9][10],andmapping[11][12].RGB-DsensorsliketheKinectarecheap,andtheextradepthinformationisinvaluableforrobotsthatinteractwitha3-Denvironment.

Recentworkongraspdetectionfocussesontheproblem

arXiv:1412.3128v2 [cs.RO] 28 Feb 2015

相关文档
...Sensitive Convolutional Neural Networks
Contour Detection Using Cost-Sensitive Convolutional Neural Networks_互联网_IT... and the softmax learning rate as ten times of the overall learning rate....
3D Convolutional Neural Networks for Human Action R...
3D Convolutional Neural Networks for Human Action...In speech and handwriting recognition, time-delay ...Based on the detection and tracking results, a ...
Convolutional Neural Networks for Sentence
Convolutional Neural Networks for Sentence_互联网_IT/计算机_专业资料。卷积神经...At test time, do not use dropout, but scale the weights by p . Like ...
...Recognition using Convolutional Neural Network(I...
Bangla Handwritten Character Recognition using Convolutional Neural Network(IJIGSP-V7-N8-5)_电子/电路_工程科技_专业资料。I.J. Image, Graphics and Signal ...
A_convolutional_neural_network_approach_for_objecti...
A neural network-based... 7页 1下载券 A ... Real-time video qualit... 暂无评价 7页 免费 ...; ; A Convolutional Neural Network Approach for ...
Accelerating Deep Convolutional Neural Networks Usi...
Accelerating Deep Convolutional Neural Networks Using Specialized Hardware_电子/...run-time (without requiring hardware re-compilation), (2) an efficient ...
UFLDL tutorials - Convolutional Neural Network
UFLDL tutorials - Convolutional Neural Network_建筑/土木_工程科技_专业资料。...(http://ufldl.stanford.edu/tutorial/supervised/FeatureExtractionUsingConvolution...
...by Hybrid Deep Convolutional Neural Networks
Convolutional Neural Networks_电子/电路_工程科技_...Later, many vehicle detection methods using ...J. Jones, “Robust real-time face detection,”...
CVPR2014视觉跟踪
Real-Time Visual Tracking Soren Hauberg*, Max ...using Convolutional Neural Networks Maryland Takumi ...Detection and Tracking with Variable Rigidity ...
Faster_R-CNN_1506.01497v3_图文
1 Faster R-CNN: Towards Real-Time Object Detection with Region Proposal ...(e.g., [4]) and region-based convolutional neural networks (RCNNs) [5...
深度学习文字识别论文综述_图文
A robust hierarchical detection method for scene text based on convolutional neural networks[C]//Multimedia and Expo (ICME), 2015 IEEE International ...
Gidi_Nave_Face_Recognition 人脸识别_图文
Face Recognition: A Convolutional Neural Network ...II. Identifying particular people in real-time (...However, it was found that using separate ...
相关主题
返回顶部
热门文档