Real-time Monocular SLAM-Why Filter.pdf

《Real-time Monocular SLAM-Why Filter.pdf》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-timeMonocularSLAM:WhyFilter?HaukeStrasdat,J.M.M.MontielandAndrewJ.DavisonAbstractWhilethemostaccuratesolutiontooff-linestruc-x0x1x2x3x0x1x2x3turefrommotion(SFM)problemsisundoubtedlytoextractasmuchcorrespondenceinformationaspossibleandperformz1z2z3z4z5z6z7z8z9z10z11z12z13z14z15z16glo

2、baloptimisation,sequentialmethodssuitableforlivevideostreamsmustapproximatethistofitwithinfixedcomputationaly1y2y3y4y5y6y1y2y3y4y5y6bounds.Twoquitedifferentapproachestoreal-timeSFMalsocalledmonocularSLAM(SimultaneousLocalisation(a)BayesianNetwork(b)MarkovRandomFieldandMapping)haveprovensuc

3、cessful,buttheysparsifytheproblemindifferentways.Filteringmethodsmarginalisex0x1x2x3x0x1x2x3outpastposesandsummarisetheinformationgainedovertimewithaprobabilitydistribution.Keyframemethodsretaintheoptimisationapproachofglobalbundleadjustment,butcomputationallymustselectonlyasmallnumberof

4、pastframestoprocess.y1y2y3y4y5y6y1y2y3y4y5y6Inthispaperweperformthefirstrigorousanalysisof(c)Filter(d)KeyframeBAtherelativeadvantagesoffilteringandsparseoptimisationforsequentialmonocularSLAM.AseriesofexperimentsinsimulationaswellusingarealimageSLAMsystemwereFig.1.(a)BayesiannetworkforSLAM

5、/SFM.(b)SLAM/SFMasmarkovrandomfieldwithoutrepresentingthemeasurementsexplicitly.(c)and(d)performedbymeansofcovariancepropagationandMontevisualisehowinferenceprogressedinafilterandwithkeyframe-basedCarlomethods,andcomparisonsmadeusingacombinedoptimisation.cost/accuracymeasure.Withsomewell-d

6、iscussedreservations,weconcludethatwhilefilteringmayhaveanicheinsystemswithlowprocessingresources,inmostmodernapplicationskeyframeoptimisationgivesthemostaccuracyperunitofcomputingtime.vision,whoseprincipleswerederivedfromphotogrammetry,andtheSimultaneousLocalisationandMapping(SLAM)I.INTR

7、ODUCTIONsub-fieldofmobileroboticsresearchhencethesomewhatLivemotionandstructureestimationfromasinglemovingunfortunatedualterminology.Theessentialcharacterofthesevideocamerahaspotentialapplicationsindomainssuchtwoproblems,estimatingsensormotionbymodellingtheasrobotics,weara