室内环境下结合里程计的双目视觉SLAM研究

:1006-4710(2009)04-0466-06SLAM王晓华1,2,傅卫平1,苏立1(1.,710048;2.,710048):针对视觉SLAM要解决的定位精度低和鲁棒性低的问题,提出一种基于双目视觉传感器与里程计信息的扩展卡尔曼滤波SLAM方法,应用改进的SIFT算子提取双目视觉图像的环境特征获得特征点,并构建出视觉特征地图;应用扩展卡尔曼滤波算法融合视觉信息与机器人位姿信息,完成同时定位与地图创建这种方法既可以解决单目视觉利用特殊初始化方法获取特征点信息不准确的问题,也可以避免双目视觉里程计利用图像信息恢复运动带来的计算量极大和运动估计不鲁棒的缺点仿真实验表明,在未知室内环境下,算法运行稳定,定位精度高:SLAM;双目视觉;里程计;SIFT;扩展卡尔曼滤波:TP24:AResearchonBinocularVisionSLAMwithOdometerinIndoorEnvironmentWANGXiao-hua1,2,FUWe-iping1,SULi1(1.FacultyofMechanicalandPrecisionInstrumentEngineering,Xi.anUniversityofTechnology,Xi.an710048,China;2.CollegeofElectronicandInformation,Xi.anPolytechnicUniversity,Xi.an710048,China)Abstract:WiththeaimofsolvingthelowpositioningaccuracyandlowrobustnessproblemsofvisionSLAMalgorithm,ExtendedKalmanFilter(EKF)methodbasedonbinocularvisionandodometerispro-posedinthispaper.FeaturepointcanbeobtainedbyextractingimagefeatureswithimprovedSIFTalgo-rithm,andthevisionfeaturemapisconstituted.SLAMiscompletedbyusingtheinformationofbinocularvisionandrobotpositionwithEKF.Thismethodcaneithersolvethemonocularvisioninaccuracyprob-lemoffeaturepointinformationobtainedbyspecialinitializationmethodoravoidtheenormouscomputa-tionbroughtaboutbybinocularvisionodometerusingimageinformationtorestoremovementaswellasthein-robustdisadvantagesofmotionestimation.Theresultsfromsimulationexperimentsindicatethatintheunknownindoorenvironments,thisalgorithmoperationisstable,andthepositioningaccuracyishigh.Keywords:SLAM;binocularvision;odometer;SIFT;ExtendedKalmanFilter(EKF):2009-08-24:(10872160):(1972-),,,,,E-mai:[e-mail](SimultaneousLocalizationandMapping,SLAM)[1,2]SLAM,,[3],SLAM,SLAM[4]SLAM[5]SLAM,CCD,[6,7],SLAMSLAM[8,9],,,466JournalofXi.anUniversityofTechnology(2009)Vo.l25No.4,,SLAM,;SIFT,,,,1SLAMSLAM,11Fig.1Systemframework,SIFT,,;,,,,(EKF),SIFT(),,,2SIFTSIFT,SIFT[10]:128,SIFT128,,,128,,,,,,SIFT,1)1,1281,11,;2)SIFT128,8,8,45b,0b;45b,,,467:SLAM,,128,,2SIFT0.4,,98236,2b1280@1024(2a)50%,2SIFTFig.2SIFTmatchingresults3SLAM,3.1,,33Fig.3Robotmotionmodel:Xr(k+1)=f(Xr(k),u(k))+Xr(k)(1),Xr(k)IRnk,u(k),Xr(k),Q(k)u(k)=[$Dk,$Hk],:xr(k+1)yr(k+1)Hr(k+1)=xr(k)+$Dk$Hk[cos(Hr(k)+$Hk)-cosHr(k)]yr(k)+$Dk$Hk[sin(Hr(k)+$Hk)-sinHr(k)]Hr(k)+$Hk+Xxr(k)Xyr(k)XHr(k)(2),xr(k)yr(k)Hr(k)k,$Dkk,$Hkk3.2,44Fig.4SensorobservationmodelZ(k)k:Z(k)=h(X(k))+E(k)(3)h,H,E(k),R(k):Z(k)=xr+xrLcosHr-yrLsinHryr+xrLsinHr+yrLcosHr+E(k)(4)468(2009)254,(xrL,yrL),Hr3.3(EKF)Taylor,EKF,,:X(k)=Xr(k)XL(k)(5),Xr(k)=[xr,yr,Hr]T,XL(k)=[X1(k),X2(k),,,Xn(k)]Tn51:ku(k)k+1ŽX(k+1)‰P(k+1):ŽX(k+1)=f(X^(k)+u(k))(6)‰P(k+1)=¨F(k)P(k)¨FT(k)+Q(k)(7),¨F(k)f(k)X(k)2:Y(k+1)3:Z(k+1)4:,+(k)K(k+1)+(k)=Y(k+1)-Z(k+1)(8)K(k+1)=‰P(k+1)HT(k+1)(H(k+1)‰P(k+1)HT(k+1)+R(k+1))-1(9)5:X^(k+1)=‰X(k+1)+K(k+1)[Y(k+1)-Z(k+1)](10)P(k+1)=(I-K(k+1)H(k+1))P(k)(11)4,,,EKF,5s,0.1m/sXr(0)=[-30,-30,0],SIFT,,ŽX(0);ŽX(1),,SIFT,Y(1),,Z(1),K(1)();,Y(1)Z(1)ŽX(1),X^(1),,EKF-5SIFT/*0,/#0,()EKF(),,,5SLAMFig.5SLAMexperimentalresults469:SLAM5,SIFT,;,,650,XY0.07m,[7]SLAM0.1m,SLAM,,76Fig.6Erroranalysis7SLAMFig.7SLAMexperimentalresultsindifferentconditions()EKF(),,7,,,;,,;,470(2009)2545SLAM,SIFT,,SLAM:[1]SmithR,SelfM,ChessemanP.Ontherepresentationandestimationofspatialuncertainty[J].InternationalJournalofRoboticsResarch,1986,5(4):56-58.[2]Durrant-WhyteH,BaileyT.Simultaneouslocalizationandmapping.PartI[J].IEEERoboticsandAutomationMaga-zine,2006,13(3):99-108.[3],(WangYao-nan,YuHong-shan).(Areviewofsimultaneouslocalizationandmapbuildingalgo-rithmsformobilerobotsinunknownenvironment)[J].(ControlTheory&Applications),2008,25(1):57-64.[4]DavisonAJ.Rea-ltimesimultaneouslocalicationandmap-pingwithasinglecamera[J].IEEETransactiononPa-tternAnalysisandMachineIntelligence,2002,24(7):865-880.[5]KimGH,KimJS,HongKS.Vision-BasedSimultaneousLocalizationandMappingwithTwoCameras:2005IEEE/RSJInternationalConferenceonIntelligentRobotsandSys-tems[C].Tokyo:IEEEPress,2005:3401-3405.[6]OrtegaJS,LemaireT,DevyM,eta.lAMoninDelayedvsUndelayedLandmarkInitializationforBearingonlySLAM:ProceedingoftheIEEEInternationalOnferenceonRoboticsandAutomationWorkshoponSLAM[C].Japan:IEEEPress,2005:1-3.[7],,(WangPeng-lin,ShiShou-dong,HongXiao-wei).SLAM(ASLAMalgorithmbasedonmonocularvisionando-dometer)[J].(ComputerSimulation),2008,25(10):172-175.[8],,(WuGong-we,iZhouWen-hu,iGuWe-ikang).(Dis-parityspacebasedbinocularvisualodometry)[J].(ChineseJournalofSensorsandActuator),2007,20(6):1432-1436.[9]DavisonAJ,NobuyukiK.3DSimultaneousLocalizationandMapBuildingUsingActiveVisionforaRobotMovingonUndulatingTerrain:ProceedingsoftheIEEEInterna-tionalConferenceonComputerVisionandRecognization[C].Hawai:lIEEEPress,2001:384-391.[10]LoweD.Distinctiveimagefeaturesfromscale-invariantkeypoints[J].InternationalJournalofComputerVision,2004,60(2):91-110.()471:SLAM