Using 3D Range Images
PANDURANGA RAO DEVARAKOTA
Do toralThesis inSignal Pro essing
TRITA-EE 2008-003
ISSN 1653-5146
ISBN 978-91-7178-850-4
SignalPro essingLaboratory
SE-10044Sto kholm
SWEDEN
AkademiskavhandlingsommedtillståndavKunglTekniskahögskolanframlägges
tilloentliggranskningföravläggandeavteknologiedoktorsexamenisignalbehan-
dlingmåndagenden28januari2008klo kan14:00iamphitheatre,BS004,Campus
Limpertsberg,UniversityofLuxembourg,Luxembourg.
©PanduRangaRaoDevarakota,januari2008
Abstra t
Thisthesisdealswiththeproblemof lassifyingautomotivevehi leo u-
pantsandestimatingtheirposition. Thisinformationis riti alindesigning
futuresmart airbag systemsproviding maximumprote tion for passengers.
A ordingtothe Ameri anNationalHighwayTra Safety Administration
(NHTSA),sin e1990,intheUSA,227deathshavebeenattributedtoairbags
deployedinlow-speed rasheswhi hin luded119 hildren,and22infants. In
these ases,intelligent deployment oftheairbag, basedonthe typeand po-
sition of o upant ould have avoided these fatalitie s. Current ommer ial
lassi ation systemsbased ontraditionalsensors, su h as pressure sensors
arenotabletodete tthepositionofo upants. Vision-basedsystemsaread-
vantageousoverpressuresensorbasedsystems,asthey anprovideadditional
fun tionalitieslike dynami o upantpositionanalysis or hildseatorienta-
tiondete tion. Onthe other hand,vision-based systemshave to opewith
several hallenges, su has, illumination onditions, temperature,humidity,
largevariationofs enes, ost,and omputationalaspe ts.
This thesis presents new pattern re ognition te hniques for lassifying,
lo aliz ing and tra king vehi le o upants using a low-resolution 3-D opti-
altime-of-ight range amera. Thissensor is apable ofprovidingdire tly
adense rangeimage,independentof theillumination onditionsand obje t
textures.Basedonthiste hnology,IEES.A.ispresentlydevelopinga amera
systemfortheappli ationofo upant lassi ation. Aprototypeofthis am-
erahasbeenthebasisforthisstudy. Therstpartofthethesispresentsthe
problemof o upant lassi ation. Herein, weinvestigate geometri feature
extra tionmethodstodis riminatebetweendierento upanttypes. Wede-
velopfeaturesthatareinvariantunderrotationandtranslation. Amethodfor
redu ingthe sizeof thefeaturesetisanalyzedwithemphasisonrobustness
and low omputational omplexity while maintaining highly dis riminative
information. Inaddition, several lassi ation methods are studied in lud-
ingBayesquadrati lassier,GaussianMixtureModel(GMM) lassierand
polynomial lassier. Weproposetheuseofa lusterbasedlinearregressio n
lassierusingapolynomialkernelwhi hisparti ularlywellsuitedto oping
withlargevariationswithinea h lass. Fulls aleexperimentshavebeen on-
du tedwhi hdemonstratethata lassi ationreliabilityofalmost100% an
bea hievedwiththeredu edfeaturesetin ombinationwitha lusterbased
lassier.
Inthis safety riti al appli ation, it isequally important to addressthe
problemofreliabilityestimationforthesystem. State-of-the-artmethodsto
estimatethereliability ofthe lassi ationare basedeitheron lassi ation
outputorbasedondensityestimation. These ondpart ofthethesis treats
estimation of the reliability of the pattern lassi ation system. Herein, a
novelreliabilitymeasureisproposedfor lassi ationoutputwhi htakesinto
a ountthelo aldensityoftrainingdata. Experimentsverifythatthisreli-
abilitymeasureoutperformsstate-of-the-artmethodsinmany ases.
Lastly,theproblemofdynami allydete tingout-of-positiono upantsis
addressed in the thirdpart of the thesis. This task requires dete ting and
methods,su h asdete ting head-likeobje ts intheimage by analyzing the
lo alshapesarenotrobustwiththe urrentsensor. Manyregionsinas ene
su hastheshoulderortheelbowoftheo upant anbein orre tlydete ted
asthehead. Inorder to opewiththese hallenges, weexploit topologyin-
formationin the range image. A modied Reeb graph te hnique has been
developed thatextra ts atopologi alskeletonof the3D ontourthat is in-
variantunderrotationand translations. Resultsverifythat theReebgraph
dete tssu essfully theheadi.e., thehead always orrespondstooneofthe
end points of the skeleton. Subsequently, a data asso iation algorithm to
sele t the orre t head andidate out of the Reeb graph andidates is pre-
sented. Results show that the resulting head dete tion algorithm basedon
First,Iwouldliketore ordmysin eregratitudetoProfessorBj
o ¨
rnOttersten,mysupervisoratKTH,fortheopportunitytopursuemydo toralstudiesinhisgroup.
I am greatly indebted to him for his immense guidan e, patien e, and support
throughout my do toralresear h. His stimulating suggestionsand ideas, sharing
hisknowledge,anden ouragementmademyresear hreallyenjoyable. IfeelBj
o ¨
rnis an ideal adviser. I would like to thank Dr. Bruno Mirba h, my supervisor at
IEE,whokeptaneyeontheprogressofmyworkandwasalwaysavailablewhenI
neededhisadvi es. Iamverythankfultohimwhohelpedmetolearnandprovided
te hni al and resear h insight to the proje t. Parti ularly, I am very grateful to
himfor hisstimulations and ontributions to theresear h. Bj
¨ o
rn andBruno, thehelpre eivedfromyoubothprofessionallyandpersonallywasimmense,andI am
gratefulto youboth ineverypossiblewayandhopeto keepupour ollaboration
inthefuture.
Thisproje twas arriedoutintheframeworkofLIASIT,UniversityofLuxem-
bourg, and in ollaboration withIEE S.A.of Luxembourg. I would liketo thank
allthe people at LIASIT for their support in every possibleway throughout the
proje t. I onveymy a knowledgement to MagaliMartin, LIASIT,for the indis-
pensable help dealing with travel funds, administration during my stay and my
ommutebetweenSto kholm,soI ouldoptimally arryoutmyresear handtrav-
els. Thanksto JoelGrotzformanydis ussionsduringthe oursework,anditwas
greattoworkwithyou. Aspe ialthankstoKarinDeminandAnnikaAugustsson,
KTH,formaking alladministrativeissuesrunningverysmoothly. Colle tiveand
individual a knowledgments are also owed to my olleagues at signal pro essing
group,KTH.
I wouldliketothankallmy oauthorsatIEE;Marta Castillo-Fran o,Dr. Ro-
mualdGinhoux,SergeKater,withouttheir onstantsupportduringdierentphases
oftheproje t,inparti ular,intheexperimentaldesignanddata olle tion,aswell
asintheimplementationofprepro essingandfeatureextra tionroutines,thiswork
would not have been ompleted in time. I would like to thankAlain Garand for
his ex ellent support during the re ordings. I also want to thank Benoit Ries,
Dr.Frederi Grandidier,ThomasSoligna fortheirlivelyworkingatmosphereand
interestingdis ussions. The time I spent with youall at IEE was rewardingand
unforgettable. I amverygrateful toDr. AloyseS hoos,IEE, whois instrumental
for LIASITproje tsatIEE.I would liketothankDr. LaurentFederspiel, Ra hel
Papa,andthepeopleintheHumanResour edepartmentatIEE,parti ularly,Mr.
Mi hael Hartmann, Cindy Hubert, and Alexandra Pereirafor their onstantsup-
portandhelpdealingwithadministrationandbureau rati mattersduringmystay
in Luxembourg.
Iwouldliketothankmythesisexamination ommitteemembers;ProfessorArne
Leijon fromKTH,Dr. MartinFritzs hefromDaimlerChryslerResear handTe h-
nology, Ulm, Professor ThomasEngel from University of Luxembourg, Professor
ChristophMeinelfromUniversityofPotsdam,Germany,andProfessorDavidBasin
from ETH, Zuri h. Myspe ial thanksto Professor HamidKrim, North Carolina
State University, the United States, for introdu ing me to Reeb graphte hnique
whi histhemaintopi ofPartIIIofthethesis,andforhistimetoa tasopponent
forthisthesis.
I am very thankful to my friends Dr. Rajesh L.V.V.L and Dr. Phani K
Yalavarthy for the wonderful time we spent on Skype, for their advi es and for
sharing many personal things. Words fail me to express my appre iation to my
wife Rekha whose dedi ation, patien e,loveand persistent onden eon me, has
takentheloadomyshoulder. I oweherforbeingunselshlyletherintelligen e,
passionsandambitions ollidewithmine. Finally,Iwouldliketothankmyparents
withoutwhomnoneofthiswouldhavebeenpossible. Theyareunbelievableinspi-
rationtome. Ithankthemforinstillingtheloveoflearninginmeandfortea hing
meaboutintegrity,dignity,andrespe t. Thisthesisisdedi atedtomyparents.
Pandu RangaRaoDevarakota
Contents vii
1 Introdu tion 1
1.1 Ba kground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 ExistingTe hnologiesforO upantSensing . . . . . . . . . . . . . . 3
1.3 A Novel3DRangeCamera . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Obje tives,Challenges,andMethodology . . . . . . . . . . . . . . . 10
1.5 TheExperimental Approa h: Measurements. . . . . . . . . . . . . . 17
1.6 OutlineandContributions . . . . . . . . . . . . . . . . . . . . . . . . 19
I O upant Classi ation 27 2 Introdu tion to O upant Classi ation 29 2.1 Introdu tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2 RelatedWork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.3 OrganizationofPartI . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3 Prepro essing 33 4 Feature Computation 39 4.1 An OverviewofFeatureExtra tionMethods . . . . . . . . . . . . . 39
4.2 FeatureExtra tionforRangeImages. . . . . . . . . . . . . . . . . . 40
4.3 Prin iple ComponentAnalysis(Eigenimages) . . . . . . . . . . . . . 45
4.4 FeatureSubsetSele tion . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.A CurvaturesoftheParaboloid . . . . . . . . . . . . . . . . . . . . . . 46
4.B Des riptionofAllFeatures . . . . . . . . . . . . . . . . . . . . . . . 47
5 Classi ation Methods 51 5.1 Introdu tion. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
5.2 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3 BayesQuadrati Classier . . . . . . . . . . . . . . . . . . . . . . . . 52
5.4 GMMClassier . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.5 PolynomialClassier (PC). . . . . . . . . . . . . . . . . . . . . . . . 53
5.6 PolynomialClusterClassier(PCC) . . . . . . . . . . . . . . . . . . 55
6 ExperimentalResults 57 6.1 Database . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
6.2 ValidationCriterionandPerforman eMeasure . . . . . . . . . . . . 58
6.3 Experiments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
6.4 Results. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
II Reliability Estimation of Pattern Classi ation Methods 67 7 Introdu tion to Reliability Estimation 69 7.1 Ba kground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.2 OrganizationofPartII. . . . . . . . . . . . . . . . . . . . . . . . . . 71
8 Data Model 73 8.1 TheLo alDensityofTrainingPatternsAroundaTest Pattern . . . 73
8.2 ApproximationUsingGaussianMixtureModel . . . . . . . . . . . . 75
8.3 Un ertaintyBasedontheLo alDensityofTrainingData . . . . . . 76
8.4 TheChoi eofParameters
C
andr
. . . . . . . . . . . . . . . . . . . 788.5 Reje tMethods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
8.A Derivationof Lo alDensityoftheTrainingPatterns . . . . . . . . . 80
9 ExperimentalResults 83 9.1 ATwo- lassSyntheti Example . . . . . . . . . . . . . . . . . . . . . 83
9.2 ExperimentsonReal-dataSets . . . . . . . . . . . . . . . . . . . . . 86
9.3 ExperimentsonO upantDataSet. . . . . . . . . . . . . . . . . . . 93
9.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
III 3D Skeleton-based Head Dete tion and Tra king 97 10 Introdu tion to Head Dete tion 99 10.1 Ba kground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
10.2 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
10.3 Dete tionApproa h . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
10.4 OrganizationofPartIII . . . . . . . . . . . . . . . . . . . . . . . . . 102
11 TopologyCodingof 3-DObje ts 105 11.1 Ba kground . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
11.2 MorseTheory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
12MethodforHead Dete tion from the Reeb Graph 109
12.1 ReebGraphExtra tiononRangeData . . . . . . . . . . . . . . . . 109
12.2 VoxelNeighborhood . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
12.3 TheReebGraphExtra tionAlgorithmforHeadDete tion . . . . . 112
12.4 TheMultipleHypothesisTra king . . . . . . . . . . . . . . . . . . . 116
13ExperimentalResults 119 13.1 EvaluationMethodologyandDatabase. . . . . . . . . . . . . . . . . 119
13.2 TopologyofExtra tedReebGraphs . . . . . . . . . . . . . . . . . . 120
13.3 GroundTruthInformation. . . . . . . . . . . . . . . . . . . . . . . . 124
13.4 EvaluationofHeadDete tionMethod . . . . . . . . . . . . . . . . . 128
13.5 EvaluationofMultipleHypothesisHeadTra kingAlgorithm . . . . 131
13.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
IV Epilogue 137
14ThesisCon lusions 139
Introdu tion
This thesis des ribes new re ognition te hniques for lassifying, lo alizing, and
tra kingtheautomotivevehi leo upantsusingalow-resolution3Dopti al time-
of-ightrange amerafordesigningso- alledsmart airbags. This hapterprovides
anoverview of urrentairbagte hnologyand pointsto theriskof airbagindu ed
injurybe auseofunne essarydeployment. Existingte hnologiestoprovidevehi le
o upant lassi ationandlo alizationaredis ussed. Theformulationofthethesis
andthesummaryof ontributionsarepresented. The hapteralsogivesanoutline
ofthethesis,andpresentsthepubli ationsforegoingthisthesis.
1.1 Ba kground
Passengersafety has be ome one of the mostimportant topi s in modern ar in-
dustry. Inthis dire tion,theadditionof airbagsto passengerrestraint systemsis
arelativelynew initiative. Airbags aredesigned as a supplement to seat-beltre-
straintsystemstoensurepassengersafetyin aseoffrontal ara idents. Airbags
wereinventedin1953andtheirdevelopmentbeganinthelate1980's. Intheauto-
motiveindustry, thein orporation ofairbagsin the presentdesign was startedin
the1990's. In aseof afrontal ar rash,airbagsbeginto inateanddeployvery
qui kly (see Figure 1.1). Thereby, airbagsredu e the han e that the o upant's
headandupperbodywillstrikesomepartofthevehi le'sinterior. Whiletheseat
beltsaretheprimaryrestraints,airbagsoersupplementaryprote tionandredu e
theriskofseriousheadinjury. Areportpublishedin2004showsthat,throughout
1987-2003,airbagssavedapproximately16,905livesintheUnitedStates[Adm04℄.
Thisstudy has shown that urrent airbagshavebeenhighly ee tive in redu ing
overallfatalities.
There have been instan es, where the o upants an also beseriouslyinjured
or killed notbe auseof afrontal rash but due to the(unneeded and unwanted)
deploymentofairbags. Forinstan e,whenanairbagdeploysonarear-fa inginfant
seat,it an ausesevereor evenfatalinjuriestotheinfant. Thesameisthe ase,
Figure1.1: Airbagdeploymentonthefrontseatpassenger.
if a passenger is too lose to the airbag while it is inating. A ording to the
Ameri an National Highway Tra Safety Administration (NHTSA), sin e1990,
in theUnited States,227deaths havebeenattributed to airbagsdeployed inlow-
speed rasheswhi hin luded119 hildrenbetweentheages1and11,and22infants
[Adm04℄.
To this end, the demand for the saferdeploymentof airbags is a knowledged
withthein reasinglyfrequentappli ationofairbagsinautomobiles. Carmanufa -
turersandvehi lesafetyadministrationsputin reasingweightonthedevelopment
of intelligent passengersafety systems, in order to redu e the in iden e of airbag
indu ed injury. A ording to FMVSS208, a safety standard was introdu ed by
NHTSA, and hasbeenfully ee tive sin e2006. A ordingto this standard, the
urrent airbag te hnology requires an o upant lassi ation system, that dea -
tivates an airbag in ase of an infant seat. The same should be the ase for a
non-o upied seat for redu ing repair osts. In addition to these safety require-
ments, it would be bene ial if thesystem oulddetermine the orientation of an
infantseatandthepositionof anadult,as this ouldallowtheadjustmentof the
airbagdeploymenttoboththetypeandposition oftheo upant. Theautomati
adjustmentofthedeploymentoftheairbagbasedono upanttypeandposition,
Figure 1.2: A pressuresensitivematrix sensor OC ® sensor) developedbyIEE for
o upantsensing;Copyright©IEE.
1.2 Existing Te hnologies for O upant Sensing
Pressure Sensitive Sensors
Therstsystemthat enteredthemarketin 2003, omplyingwiththenewregula-
tionsforo upant lassi ation,usesseatbasedsensorsthat lassifytheo upan y
a ordingtoweight. OneofthesesystemsistheO upantClassi ationsensorsys-
temOC ®sensordevelopedbyIEES.A.([IEE℄). Theo upant lassi ationsystem
onsistsofapressuresensitivematrixsensor (seeFigure1.2),whi hmeasures the
two-dimensionalpressureprolein theseatarea. A patternre ognitionalgorithm
isusedasabasisforo upant lassi ationdepending onthemeasuredpressured
prole. Thissystemallowsthesuppressionofanairbag inthepresen eofa hild
withor withouta hildrestraintsystem,while allowingafull airbagdeployment
in the presen e of anadult person. However,in ontrast to the existing airbag
system,futuresystemswillbemoreadvan edredu ingtheriskandmagnitude of
injuriesbyautomati allyadaptingtheairbagandseatbeltpretentionto thedriv-
ingstatusofthevehi le,itso upants,and the rashseverity. Su h sophisti ated
airbagdeploymentstrategiesrequireinformationaboutthesizeoftheo upantand
his/herposition,relativeto theairbag module in order to provideea h o upant
withoptimalprote tionandsafety. Theexistingte hnologiesbasedonthepressure