Real-Time Grasp Detection Using Convolutional Neural Networks

《Real-Time Grasp Detection Using Convolutional Neural Networks》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、Real-TimeGraspDetectionUsingConvolutionalNeuralNetworksJosephRedmon1,AneliaAngelova2Abstract—Wepresentanaccurate,real-timeapproachtoroboticgraspdetectionbasedonconvolutionalneuralnetworks.Ournetworkperformssingle-stageregressiontograspableboundingboxeswithoutusingstand

2、ardslidingwindoworregionproposaltechniques.Themodeloutperformsstate-of-the-artapproachesby14percentagepointsandrunsat13framespersecondonaGPU.Ournetworkcansimultaneouslyperformclassificationsothatinasinglestepitrecognizestheobjectandfindsagoodgrasprectangle.Amodificationto

3、thismodelpredictsmultiplegraspsperobjectbyusingalocallyconstrainedpredictionmechanism.Thelocallyconstrainedmodelperformssignificantlybetter,especiallyonobjectsthatcanbegraspedinavarietyofways.I.INTRODUCTIONPerception—usingthesenses(orsensorsifyouarearobot)tounderstandyo

4、urenvironment—ishard.Visualper-ceptioninvolvesmappingpixelvaluesandlightinformationontoamodeloftheuniversetoinferyoursurroundings.Gen-eralsceneunderstandingrequirescomplexvisualtaskssuchassegmentingasceneintocomponentparts,recognizingwhatthosepartsare,anddisambiguating

5、betweenvisuallysimilarobjects.Duetothesecomplexities,visualperceptionisalargebottleneckinrealroboticsystems.Generalpurposerobotsneedtheabilitytointeractwithandmanipulateobjectsinthephysicalworld.HumansseeFig.1.TheCornellGraspingDatasetcontainsavarietyofobjects,eachnove

6、lobjectsandknowimmediately,almostinstinctively,withmultiplelabelledgrasps.Graspsaregivenasorientedrectanglesinhowtheywouldgrabthemtopickthemup.Roboticgrasp2-D.detectionlagsfarbehindhumanperformance.WefocusontheproblemoffindingagoodgraspgivenanRGB-DviewInsteadoflookingon

7、lyatlocalpatchesournetworkusesoftheobject.globalinformationintheimagetoinformitsgrasppredic-WeevaluateontheCornellGraspDetectionDataset,antions,makingitsignificantlymoreaccurate.Ournetworkextensivedatasetwithnumerousobjectsandground-truthachieves88percentaccuracyandruns

8、atreal-timespeedslabelledgrasps(seeFigure1).Recentworkonthisdataset(13framespersecond).Thisredefinesthestate-of-the-ar