Salient Regions Detection for Indoor Robots using RGB-D Data(3)

Weexpecttodetermineasaliencymapfromtheseregions.Basedonsegmentedregions,thesaliencymeasurementsarecalculatedbyforegroundsaliencyfeaturesandbackgroundinformation,duetouniquenessoftheforegroundandcon-sistencyofthebackground.

Weassumethatforegroundsalienceissigni cantanddistinctiveinanimage.Thedistinctivenessreferstoaregionwithhighdifferencefromitsneighborhood.Inthissection,weproposeanewforegroundsaliencemeasurebasedondistributionofcolor,depth,positionandareaasanextensiontotheapproachby[3].

Inanimage,theforegroundsaliencevalueFSofeachsegmentedregioniscalculatedas:FS(ri)=ar(ri)·

N j=1

cd(ri) cd(rj) 2·wfs(ri,rj)

Fig.2.Segmentationexamples.RawRGBandgray-scaledepthsamples

(toptworows),graph-baseddepthsegmentation(thethirdrow),graph-basedcolorsegmentation(thefourthrow)andgraph-basedRGB-Dsegmentation(bottomrow)

(2)

ar(ri)isthearearatiooftheregionritotheentireimageandisusedasanadditionalweightingfactorasopposedto[3].Nrepresentsthenumberofsegmentedregions,andcd(r)againisextendedtousedepthinformationandrepresentsaveragecolorinCIELab-spacecombinedwithdepthinregionr.TheGaussianweightwfs(ri,rj)isalocalcontrasttermofforeground,asintroducedin[3]:

wfs(ri,rj)=exp(

1

pi pj 2)δfs

(vi,vj)∈Ehasaweightvaluew(vi,vj)tomeasurethe

dissimilaritybetweenneighbouringverticesviandvj.Thesimilaritybetweenverticesinhomogeneousregionsishigherthanindiscrepantregions.Inourcase,theweightw(pi,pj)referstothedistancebetweenadjacentpixelsinRGBandthegray-scaledepthmapsimplyde nedas:

w(pi,pj)= RGBD(pi) RGBD(pj) (1)

2

= RGB(pi) RGB(pj) +(D(pi) D(pj))2

FS(ri)effectivelyrepresentstherarityofaregionriwithcolor,depthcdiandareacomparedtoallotherregionsrj.C.BackgroundMeasure

Tobeabletoef ciently lterouttruenegativeaswellasfalsepositivesaliencycandidates,itisimportanttode neadequatemeasuresthatareabletoidentifytherespectiveregions.Sinceimagebackgroundusuallyfeaturessomenicepropertieslikehigharearatioandawidespreadclosetotheimageborders,thosecharacteristicsareoftenmodelledinliteratureforsaliency ltering.Whilesalientregionsmayberegardedaslocalregioncandidates,withahighvarianceinsmallareas,backgroundregionsmaybeinterpretedastheircounterpartfeaturingglobalhomogeneityandhighdepthvalues,andasthusbeinghighlysuitableforbroad-phase ltering.Wefollowthiswellknownparadigmandwishto ndabackgroundrepresentationthat,combinedwiththeforegroundrepresentation,isabletoremoveorweakenfalsesaliencycandidatesandthusdeliverbetterestimates.

Takingtheseconsiderationsintoaccount,weproposeameasureCDis(ri)toquantifythepositionofaregionriinanimage.Itisde nedas

CDis(ri)=1

minj p¯i,cj

dc

(3)

Here,RGB(pi)representsthe3-dimensionalvectorofred,greenandbluevaluesofthepixelpiinRGBcolorspace.Respectively,D(pi)isthegray-scaledepthrepresentingthedistanceofthepixelpi.We rstlynormalizefourchannelstotherange[0...255].Toreducein uenceofnoiseartifacts,wethenapplyaGaussian ltertosmootheachofthefourchannelsbeforecalculatingtheedge-weights.Finally,weconstructaminimumspanningtree(MST)tomergesimilarregionsusingtheminimumweightededgebetweentheregions.

Fig.2showsthegraph-basedsegmentationresultsusingtheRGB,depthandthecombinedcoloranddepth-basedsegmentation,respectively.ResultsindicatethatfusionofRGBanddepthprovidesthebestsegmentationquality.Thisisreasonablesincecolorisaninformativebutsensitivefeature,whilethedepthvalueisverywellabletocapturethecompactnessofobjects,andthusmaybeusedtoremoveover-segmentationusingsolelycolor.B.ForegroundSaliencyMeasure

Aftersegmentation,weobtainmultipledifferentregions,eachofwhichhassimilarhomogeneousinnerproperties.

wherep¯irepresentsthecenterpositionofregionri,cjrepresentsthecornerpositionsoftheimageanddcisthedistanceoftheimagecentertoitscorners.Assuming,thatbackgroundisfarawayfromtheimagecenter,CDis(ri)maybeinterpretedgeometricallyasthedistancefromone