简介:Thispaperpresentsahandheld3Dvision-basedscannerforsmallobjectsbyusingKinect.Itisdifferentfromthepreviouscolor-glove-basedapproacheswhichrequiresegmentingthetargetobject.First,weeliminatethenoisesandtheoutlierscausedbyholdinghands.Second,weapplyKinect-fusionalgorithmandtruncatedsigneddistancefunction(TSDF)torepresent3Dsurfaces.Third,weproposeamodifiedintegrationstrategytoeliminatethehandeffect.Fourth,wetakeadvantageoftheparallelcomputationofGPUsforreal-timeoperation.Themajorcontributionsofthispaperare(1)theregistrationprecisionisimproved,(2)theofflineamendmentandloopclosureoperationarenotrequired,and(3)concave3Dobjectreconstructionisfeasible.IndexTermsHandheld3Dscanning,Kinect-fusion,Truncatedsigneddistancefunction(TSDF).1.IntroductionRecently,thesensor-based3Dmodelreconstructionmethodshavebeenproposed[1].Thesensordeviceshavedifferentpropertiessothatthe3Dreconstructionalgorithmsvaryaccordingly.Thecommonlyusedsensordevicesaretime-of-flight(ToF)cameras[2]-[4],laserscanners[5],andstructuredlightscanners[6],[7].Lasershavegainedareputationforaccuracy;however,caremustbetakentouseeye-safelaserswhenoperatinginproximitytohumans.Foraninteractivesystem,thestructuredlightscannerwhichisbasicallyapassivevision-basedsensordeviceissuperiorbecauseitprovidesa2DdepthimageperframeandismoreaccuratethanthatofaToFcamera.Here,wepresentareal-time3DscannerusingthedepthimagescapturedbyKinect.