• Aucun résultat trouvé

Constructing the Exact Voronoi Diagram of Arbitrary Lines in Space

N/A
N/A
Protected

Academic year: 2023

Partager "Constructing the Exact Voronoi Diagram of Arbitrary Lines in Space"

Copied!
20
0
0

Texte intégral

(1)

HAL Id: inria-00480045

https://hal.inria.fr/inria-00480045

Submitted on 3 May 2010

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

Constructing the Exact Voronoi Diagram of Arbitrary Lines in Space

Michael Hemmer, Ophir Setter, Dan Halperin

To cite this version:

Michael Hemmer, Ophir Setter, Dan Halperin. Constructing the Exact Voronoi Diagram of Arbitrary

Lines in Space. [Research Report] RR-7273, INRIA. 2010, pp.19. �inria-00480045�

(2)

a p p o r t

d e r e c h e r c h e

0 2 4 9 -6 3 9 9 IS R N IN R IA /R R -- 7 2 7 3 -- F R + E N G

Algorithms, Certification, and Cryptography

Constructing the Exact Voronoi Diagram of Arbitrary Lines in Space

with Fast Point-Location

Michael Hemmer — Ophir Setter — Dan Halperin

N° 7273

May 2010

(3)
(4)

Centre de recherche INRIA Sophia Antipolis – Méditerranée 2004, route des Lucioles, BP 93, 06902 Sophia Antipolis Cedex

with FastPoint-Loation

Mihael Hemmer

, Ophir Setter

,Dan Halperin

Theme: Algorithms,Certiation,andCryptography

Algorithmis,Programming,SoftwareandArhiteture

Équipes-ProjetsGeometria

Rapportdereherhe n° 7273May201016pages

Abstrat: We introdue a new, eient, and omplete algorithm, and its

exatimplementation,to omputetheVoronoidiagram oflinesin spae. This

is a major milestonetowardsthe robust onstrution of the Voronoi diagram

of polyhedra. As we follow the exat geometri-omputation paradigm, it is

guaranteed that we always ompute the mathematially orret result. The

algorithmisompletein thesense thatitanhandleallongurations,inpar-

tiular all degenerate ones. The algorithm requires

O(n 3+ε )

time and spae,

where

n

isthenumberoflines. TheVoronoidiagram isrepresentedbyadata struture that permits answering point-loation queriesin

O(log 2 n)

expeted

time. The implementation employs the Cgal pakages for onstruting ar-

rangementsandlowerenvelopesonparametrisurfaestogetherwithadvaned

algebraitools. Supplementarymaterialandinpartiulartheprototypialode

of our implementation anbe found in the website: http://ag.s.tau.a.

il/projets/internal-projets/3d-lines-vor/projet- page.

Key-words: VoronoiDiagram,PointLoation,LowerEnvelope,ExatGeo-

metriComputing,CGAL

ThisworkhasbeensupportedinpartbytheIsraelSieneFoundation(grantno.236/06),

by the German-Israeli Foundation (grant no. 969/07), and by the Hermann Minkowski

MinervaCenterforGeometryatTelAvivUniversity.

Geometria,INRIASophiaAntipolis,Frane;Mihael.Hemmersophia.inria.fr

ShoolofComputerSiene,Tel-AvivUniversity,Israel;[ophirset,danha℄post.tau.a.il

(5)

droites quelonques dans l'espae

ave loalisation eae

Résumé: Nousprésentonsunnouvelalgorithmeeaeetomplet,ainsique

sonimplantationexate,pouralulerlediagrammedeVoronoïdedroitesdans

l'espae. C'est une étape majeureverslaonstrution robuste du diagramme

deVoronoïdepolyèdres. Noussuivonslemodèledualul géométriqueexat,

il est dongaranti quenous alulons toujours le résultat mathématiquement

orret. L'algorithmeestompletenesensqu'ilpeuttraitertouteslesong-

urations, enpartiuliertous lesas dégénérés. L'algorithmea une omplexité

O(n 3+ε )

entempsetenespae,

n

estlenombrededroites. Lediagrammede

Voronoïestreprésentéparunestruturededonnéespermettantderépondreaux

requêtesde loalisationsdepointsave uneomplexité moyenneen

O(log 2 n)

.

L'implantation utilise les modules CGAL de onstrution d'arrangements et

d'enveloppesinférieures surdes surfaesparamétriques ainsiquedes outilsal-

gébriques avanés. Lematériel supplémentaireet en partiulierle ode proto-

type de notre implantation peuvent être trouvés sur lesite: http://ag.s.

tau.a.il/projets/internal-projets/3d-lines-vor/projet- page.

Mots-lés : DiagrammedeVoronoï,Loalisation,EnveloppeInférieure,Cal-

ulGéométriqueExat,CGAL

(6)

1 Introdution

TheVoronoidiagram(

VD

)isamongthemostfundamentalstruturesinCom- putational Geometry,andisknowntobeausefultoolin avarietyof domains.

Forinstane,struturalbiology[24,35℄androbotmotionplaning[19,34℄apply

Voronoi diagrams to enode point sets keepingmaximal distane from atoms

or obstales, respetively. A related onept is the medial-axis transform [6℄,

whih is onsidered fundamental in solid modeling and applied to problems

suhasnite elementmeshing, shapemorphing,andfeature reognition. Yet,

theadaptation ofomplexthree-dimensional Voronoidiagrams in professional

tools has been very slow. Their use is hindered by the diulty of design-

ing and implementing reliablegeometri algorithms for omplexstrutures in

three-dimensionalspae.

Voronoidiagramshavebeenthesubjetofatremendousamountofresearh.

WereferthereadertothesurveybyAurenhammerandKlein[2℄ofworkpub-

lisheduptill2000. Voronoidiagramsin

R 2

arewellunderstoodinalmostallas-

pets,thatis,intermsofomplexityandoptimalalgorithmsaswellasinterms

ofrobustandeientimplementations. In

R 3

muhlessisknown,evenforsim-

ple objetssuh aslines, segments, orpolyhedra. Forexample, atight bound

on the ombinatorial omplexity of the

VD

of

n

lines or line segments in

R 3

isunknown; itis onjeturedthattheomplexityisnear-quadrati;theknown

lower bound is

Ω(n 2 )

[1℄, but the best known upper bound is1

O(n 3+ǫ )

[31℄.

Intheaseoflineswith axednumber

c

oforientationstheupperbound was improvedto

O(c 4 n 2+ε )

[25℄. A omplete analysis ofall possibleombinatorial asesforthree arbitrarylines ispresentedbyEverettetal.[17,18℄.

Today,therearemanypublishedresultsonrobustonstrutionsofdierent

typesofVoronoidiagramsin

R 2

. NotonlyVoronoidiagramsofpointsareon-

sidered,butalsoVoronoidiagramsoflinesegments[22℄,irles[15℄,ellipses[16℄,

and more [8, Ÿ2℄. In

R 3

, an exat implementation of the Voronoi diagram of additively-weightedpointswasanalyzedin[7℄,butwearenotawareofanyex-

at, omplete,and implementedalgorithm that omputesVoronoidiagrams of

lines,linesegments,orpolyhedra. Nevertheless,progresshasbeenmadetoward

theexatomputationofthearrangementofquadris[5,13℄. EahVoronoiell

ofthediagramoflinesinspaeanberepresentedastheunionofellsofsuh

an arrangement. Other approahesexpliitlyaim foran exatorrobustom-

putationof theVoronoidiagram (or themedial axis)[10,26℄. However,those

approahes are not omplete. For example, Culver'salgorithm [10℄ does not

handlesingulartrisetor-urves.

Finally, Hanniel and Elber [20℄ provided an algorithm to onstrut the

Voronoi ell of bounded planes, spheres, and ylinders in

R 3

. Though it is

in some aspets similar to ours, the approah is approximate, does not deal

withdegeneraies,andleavesrobustnessissuesaside.

Wepresentanexatandomplete(andthusrobust)algorithmforomputing

theVoronoidiagram ofarbitrarylines in threedimensionswith respetto the

Eulidean metri. The algorithmrequires

O(n 3+ε )

time andspae, where

n

is

thenumberofinputlines. Thedatastrutureadmitsansweringofpoint-loation

queriesin

O(log 2 n)

time. Weantiipate thatthe natureof thealgorithm and

1

Aboundoftheform

O(f(n) · n ε )

meansthattheatualupperboundis

C ε f(n) · n ε

,for

any

ε > 0

,where

C ε

isaonstantthatdependson

ε

,andgenerallytendstoinnityas

ε

goes

to

0

.

(7)

thegeneralapproahofitsimplementationonstituteamajormilestonetowards

anexatandrobustonstrutionoftheVoronoidiagramofpolyhedrain

R 3

.

Weutilizethe fatthat in Eulidean spae theVoronoiellanbeonsid-

eredasalowerenvelopesinetheellessentiallyhasaertainstarshapedness

property: Foranypoint

p

insidetheVoronoiellofaspeilinesite

,theline

segment onneting

p

to itsprojetion

p ℓ

onto

,is fully ontainedin theell.

ThisobservationenablesustorepresenttheVoronoiellof

asaminimization diagram,whihis(oneptually)embeddedonaninnitesimallysmallylinder

around

. Thisobservationissimilar to (butnotthesameas) thewell-known onnetionbetweenVoronoi diagramsand lowerenvelopes[14℄. Lowerdimen-

sionalellsarerepresentedseveraltimes,namelyaspartoftheboundaryofthe

VC

ofeahline theyare assoiatedwith. Theimplementationisdevelopedin andbasedonCgal,ComputationalGeometryAlgorithmsLibrary.

2

Thepaperisorganizedasfollows. Setion2disussespreliminarysubjets,

suh as properties of bisetors and trisetors of lines in spae and the lower

envelope algorithm. Setion 3 desribes the details of the onstrution of a

Voronoiell. Setion 4disussesthepointloationalgorithm anditsanalysis.

Setion 5gives implementation details and presentspreliminary experimental

resultsthatwereobtainedwithoursoftware.

2 Preliminaries

Let

O = { s 1 , s 2 , . . . , s n }

be a set of objets in

R d

, also referred to as sites.

We follow the Voronoidiagram denition by Everett et al. [18℄: TheVoronoi

diagram

VD( O )

is the subdivision of

R d

into ells, where eah ell

VC (S)

is

assoiated with a subset

S ⊆ O

, suh that every point in

VC(S)

is stritly

loser to all sites in

S

than to all other sites in

O

and is equidistant from all sitesin

S

. Theformaldenition is:

VC(S) =

p ∈ R d

∀ s ∈ S, t ∈ O \ S : d(p, s) < d(p, t)

∀ s, t ∈ S : d(p, s) = d(p, t)

In theontextof this paper,

O

denotesa set of arbitraryrational lines in

R 3

and

d( · , · )

denotestheEulideandistanefuntion. Thesetofpointsthat isof

equaldistanetotwoorthreesitesisalledabisetorortrisetor,respetively.

where

d( · , · )

denotes theEulidean distanefuntion. Whentheardinally

of

S

istwoorthreeitisalledbisetorortrisetor, respetively.

2.1 Properties of Bisetors and Trisetors

WenextstatesomepropertiesofbisetorsandtrisetorsoftheVoronoidiagram

oflinesin

R 3

thatareusedthroughoutthispaper. Proposition1givesproperties ofbisetors;seeFigure1forillustrations.

Proposition1 The bisetorof twolines

ℓ 1

and

ℓ 2

in three-dimensional spae iseither(a)ahyperboliparaboloid(asurfaeofalgebraidegree2),if

ℓ 1

and

ℓ 2

areskew,(b)aplane,if

ℓ 1

and

ℓ 2

areparallel,or()apairoforthogonalplanes,

if

ℓ 1

and

ℓ 2

areonurrent. In the latterase, the singularlousof thebisetor

2

www. gal. org

(8)

(a) (b) ()

Figure1: Bisetorof:(a)twogenerilines;(b)twoparallellines;()twointersetinglines.

Thediagramswerereatedwithourimplementation(seeSetion5),and werelippedbya

sphereforonveniene.

isalinethat isperpendiular to

ℓ 1

and

ℓ 2

andpassesthrough theirintersetion point.

Themain theoremofEverettetal. [17,18℄providesagoodoverviewofthe

dierentasesofthetrisetor:

Theorem2 (Everett etal.) The trisetor of three lines is either (i) a non-

singular quarti, if the three lines are pairwise skew but not all parallel to a

ommonplanenoronthe surfaeofahyperboloidofrevolution,(ii)aubiand

a line that do not interset, if the three lines are pairwise skewand lie on the

surfaeofahyperboloid ofrevolution,(iii)anodalquarti,if thethreelinesare

pairwiseskewandallparalleltoaommonplane,(iv)oneparabolaorhyperbola,

ifthereisexatlyonepairofoplanarlineswhihareparallel,(v)twoparabolas

or hyperbolas that interset, if there is exatly one pair of oplanar lines that

interset, (vi) between 0 and 4 lines, if there are two pairs of oplanar lines,

or (vii) one line, in the ase of three oplanar onurrent lines, the ommon

singularlousof the bisetors.

WeuseaorollaryoftheabovetheoreminSetion3,wherewedesribethe

onstrutionofaVoronoiellin thediagramoflines.

2.2 Lower Envelope Algorithm

Again, we regardeah three-dimensional

VC

asalowerenvelopewith respet

toitslinesite

ℓ 0

. Thisloverenvelopeisrepresentedasaminimization diagram whihisoneptuallyembeddedinin the

uv

-parameterspaeofthesurfaeof

aninnitesimallysmallylinderaround

ℓ 0

.3 Weutilizethedivide-and-onquer algorithmforonstrutinglowerenvelopes[1℄asitisimplementedinCgal[32,

Ÿ8.5℄,whihwebrieydesribenext.

Sinethealgorithmprojetsbisetorsintotheparameterspae,allbisetors

are initiallysplit upinto

uv

-monotonesurfaes. Thealgorithm thensplits the

resultingset

G

intotwosubsets

G 1

and

G 2

ofroughlyequalsize,andreursively

omputestheir minimization diagrams

M 1

and

M 2

. Inthe onquerstep, the

twodiagrams are mergedinto one. First,the overlayof

M 1

and

M 2

is om-

puted, whereeahfeatureislabeledwithuptotwosets oflabels

L 1

and

L 2

of

andidatesurfaes frombothdiagrams. Thereafter,thearrangementis further

renedsuhthat eahfeatureaneitherbelabeledwith

L 1

,

L 2

,or

L 1 ∪ L 2

. In

3

SeeSetion3fordetailsonthe

uv

-parameterspaesetting.

(9)

partiular, eahfae that is labeled withtwobisetorsis rened by theorre-

spondingprojetedtrisetorurve. Notethatthis stepanalsosplitupedges.

After theomparison of bisetorsthealgorithm removesredundantedges and

verties,whihyieldsthenaldiagram. Theomplexityoftheabovealgorithm

is

O(n 2+ε )

,withtheonditionthatthebisetorsurfaesarewell-behaved.

Note that the algorithm heavily relies on arrangement operations suh as

overlay,whihareprovidedby[32,Ÿ8.1℄and[4℄. Thoughwetreatthisasablak

boxthroughoutmostofthepaper,somemoredetailsanbefoundinSetion5.

Theadditionalonstrutionsandprediatesrequiredbythelowerenvelopealgo-

rithmare: theonstrutionoftheprojetedboundaryof

uv

-monotonesurfaes,

theonstrutionoftheprojetedintersetionoftwo

uv

-monotonesurfaes,and

theomparisonoftwobisetorsaboveafae,edge,orvertex.

3 Computing a Voronoi Cell

Thissetiondisussestheomputationofthe

VC

ofoneline, referredtoasthe

baseline anddenotedby

ℓ 0

.

The two-dimensional pakage of Cgal has the infrastruture to ompute

envelopesoverylinders. However,fortheeienyoftheimplementationitis

importanttokeepthealgebraidegreeoftheprojetedurvesaslowaspossible.

Therefore,weprojettheurvesontwoparallelplanesthatsandwih thebase

line,whilekeepingtheprojetiondiretionnormaltotheylinder. Thisredues

themaximumdegreeofaprojetedtrisetor urvefrom sixteendownto eight.

3.1 Parametrization and Projetion

Let

F = { − → b 1 , − →

b 2 , − →

b 3 }

beanorthogonalbasisof

R 3

whihishosensuhthat

− → b 1

is

thediretionofthebaseline

ℓ 0

. Moreover,let

p 0

besomerationalpointon

ℓ 0

.

Now,onsider the parametrization

X (u, v, r) = p 0 + u · − →

b 1 + v · r · − →

b 2 + r · − → b 3

.

Notethat

X (u, v, ± 1)

denestwoparallelplanes(

uv

-planes)thatsandwih

ℓ 0

,

whih we all the positiveand the negativeplane, respetively. Thus a point

X (u 0 , v 0 , ± 1)

representsaraythatoriginatesfrompoint

p 0 + u 0 · − →

b 1

on

ℓ 0

with

diretion

± (v 0 · − → b 2 + − → b 3 )

.

Notethattheplane

H = { x ∈ R 3 | (x − p 0 ) T · − → b 3 = 0 }

isnotoveredbythe

parametrization. However, onean simplyglue the arrangementstogether as

longasthehosenframeisgeneri,thatis,urvesarenotallowedtotouh

H

,

interset in

H

, or even be ontainedin

H

. However, urvesare allowedto

transverselyinterset

H

,whereeahintersetiongivesrisetoasimplevertial asymptoteintheprojetion.

Inordertoavoidtheseritialases,wegeneratetheloalframebysetting

→ b 2

to somerandomvetorthat is orthogonalto

→ b 1

. Thoughthisframe isgeneri

with high probability, we also hek in all relevant prediates that the frame

is indeed generi. If neessary, we restart the omputation hoosing another

random frame. We hose the standardstrategy that inreasesthe number of

randombitsused foreahiteration. Thiswayweguaranteeterminationanda

smallnumberofadditionalbitsdue totherandomization.

Wehighlightbelowseveralmajorissuesintheprojetionofatrisetor. The

projetionof abisetors' boundaryand adetailed ase analysis is deferredto

(10)

AppendixA . Werelymerelyonthegeneriframeandonthefollowingorollary

that diretlyfollowsfromTheorem 2:

Corollary 3 Thesetofpointswherethetrisetordoesnotrepresentatransver-

sal intersetion of the bisetors is a 0-dimensional set, namely, the singular

pointsof thetrisetor. Theonly exeptionisthe aseofthreeoplanaronur-

rent lines; inthis asethe trisetor isthe ommon singular lous (line) of the

three bisetors.

For atrisetor

T 0ij

let

B 0i , B 0j , B ij ∈ Q [x 1 , x 2 , x 3 ]

be the three trivariate

polynomialsoftherelevantbisetors. Nowlet

B 1

and

B 2

bethetwobisetors

of minimal degree,

d 1

and

d 2

, respetively. The projetionis arried outbya resultant omputation [33℄. However,sine wewish to projettowards

ℓ 0

we

rstsubstitute

X (u, v, r)

into

B 1

and

B 2

andomputetheresultantwithrespet

to

r

.

res(u, v) := resultant(B 1 ( X (u, v, r), B 2 ( X (u, v, r)), r) ∈ Q[u, v] .

→ v 2

→ v 3

→ v 1

B 0 i

B 0 j

X (u, v,+1)

X (u, v, − 1) v →

← v ℓ 0

T 0 ij

T 0 ij

cross-section perpendicular to ℓ 0

Thisisatmostabivariatepolynomialof

degree

2d 1 d 2

. Thus, in theworst ase (the

generi ase) this is an irreduible polyno-

mial of degree 4

only

8

. However,due to its

algebrai nature the approah an not im-

mediately distinguish between the positive

andthenegativeparameterspae. TheFig-

uretotherightillustrateshowtheresultant

projets

T 0ij

into the positiveand negative

plane. Werstsplituptheprojetedurveinto

u

-monotonesegmentsusing[4℄.

Inpartiular,urvesaresplitupatvertialasymptotes.

Inordertodeidethatanar

α

isonaertainplain weutilizeCorollary3,

namelytheobservationthatinallbutoneexeption(whihishandledexpliitly)

twobisetorsmustintersettransverselyalongthetrisetorurve,whihimplies

that theymustinterhangetheirorderwhilepassingtheprojetedtrisetor.

Thisisdetetedbytworayshootsatrationalpointsrightaboveandbelow

α

.

Let

p

and

p

bethese twopoints, respetively. Toensurethat bothpoints are hosensuientlylose,weonstrutarationalvertialline

L

thatintersets

α

initsinterior,sayatpoint

p α

. Wehoosethepointson

L

suhthattheyisolate

thearfromallother intersetionsof

L

with

res

. Nowonsiderthepathon

L

from

p

(or

p

)to

p α

.

p

issuientlyloseto

α

sinethispathdoesnotinterset

res

until itreahes

α

. Inase

α

isvertial,wehoose

L

tobehorizontal.

Furtherdetailsandin partiularhowtoguaranteethatthehosenframeis

generi,anbefoundinAppendix A.

3.2 Lower Envelope Prediates

A ore part of the envelope algorithm is the representation of minimization

diagrams as labeled arrangementsand the overlayof suharrangements. The

requiredprediates forthese operations relateto planaralgebraiurvesonly,

4

More preisely, itisa bivariatepolynomialofbi-degree atmost

(4, 4)

. Fora standard

rationalparametrizationoftheylinder,wewouldobtainapolynomialofbi-degree

(8, 8)

or

16

intotal.

(11)

whih are provided by[4℄. However,it remains to ensurethat nointersetion

takesplaein

H

. This boilsdownto testingthattheleadingoeientswith

respet to

v

of twonon overlapping (o-prime) urveshave no ommon root.

Thus, weprovideaslightly modiedset ofprediatesthat additionallyensure

thisondition.

Theprediatesthatareadditionallyrequiredbytheenvelopealgorithmare

theomparisonoftwobisetorsaboveafae,anedge,oravertex,respetively;

seealsoSetion2.2. Forafae,itissuienttoseletarationalpointinsideit

and omparethesurfaes along theorrespondingray. The pointis hosenin

asimilar way to theapproah used whensorting thetrisetorsto thepositive

and negative planes. For an edge we onstrut a vertex in its interior and

omparealongtheorrespondingray. Foravertex,whihmaynothaverational

oordinates, we rst hek whether the point is on the projetedintersetion

of the two bisetors, and report equality if it is indeed the ase. Otherwise

weomparebisetorsatarationalpointsuiently loseto thevertex, where

suientlyloseisagaindeterminedbyasimilarstrategyasinsortingtrisetor

urves.

3.3 Complexity

Forthetimeomplexityandspaeomplexityanalysisweignoreadditionalosts

that mayarise due to variable bit-lengthof various implementationsadhering

totheexatomputationparadigm[36℄. Wealsoignoretheadditionalrun-time

that anresultfrom apoorhoieofageneriframe(Setion3.1),asitisnot

thegeneralase,andhasnoimpatonperformaneinexpetation.

Thebisetorsurfaes arealgebraisurfaes ofmaximumdegreeof

8

there-

fore,thetime-omplexityof thelowerenvelopealgorithmis

O(n 2+ε )

(whih is

alsothebestknownupperbound). Thustherun-timeomplexityofomputing

theellsforall

n

linesis

O(n 3+ε )

,whihalsobounds thespaeomplexity.

4 Fast Point Loation

Given a query point

p

we wish to nd the losest line to it. Consider the

followingpoint-loationstrategy: Westartwitharandomline site

. Firstwe

projet

p

on

andloateitsimageintheminimizationdiagramof

. Theimage

isloatedonafeature oftheminimizationdiagramwhihislabeledwitha(in

generalnotempty)setof linesites

S

. Wethen omparethedistane

d(ℓ, p)

to

d(ℓ , p)

foroneline

∈ S

. If

d(ℓ, p)

islessthanorequalto

d(ℓ , p)

wereport

or

S ∪ ℓ

,respetively. Otherwiseweontinueintheellof

. Thiswalkthrough

theVoronoidiagramterminatessinethereisonlyanite numberofellsand

thedistaneof

p

totheurrentlinealwaysdereases. Weanloatetheimage

of

p

insidetheminimizationdiagraminexpeted

O(log n)

timebyusingpoint-

loation based on trapezoidal deomposition [27℄. Combining this algorithm

withtheideaoflandmarks[21℄mayalreadyhavegoodperformaneinpratie.

However,thealgorithmhasaworst-timetimeomplexity

O(n log n)

.

Weturnitintoanalgorithmwithatime-omplexity

O(log 2 n)

byombining

itwith astrategythat issimilar toskip lists. Webuild ahierarhyofVoronoi

diagrams. Thelowest layerontainsthe

VD

of thefull setof lines, whileeah

other layersontainthe

VD

ofonly

1/k

(random)linesof thepreedinglayer,

(12)

where

k > 1

issomeonstant. Thehighestlayer(therootlayer)ontainsonlya

onstantnumberoflines,andthenumberoflayersis

O(log n)

. Inordertoloate

apoint

p

werstloateitin the root layerusing the walkstrategy desribed

above. Wethenproeedtothenextlayerstartingatthelinethatwasfoundin

thepreedinglayer.

The following theorem summarizes the performane of the point-loation

struture(see[11,23℄ forsimilar analysisin2D):

Theorem4 The expetedrunningtimeofthe point-loation queryinthe hier-

arhial VDstrutureis

O(log 2 n)

.

Proof: The number of ells visited at the root layer is obviously at most

k

. For all otherlayers,onsider thethepath bakward,from its targetto the

soure: foreveryelltheprobabilitythatitisalreadythesoureis

1/k

. Thus,

the expeted length of a path is

P n i=1

i

k ( k k 1 ) i 1 ≤ k

. That is, theexpeted

running-timeis

k P log k n

i=1 T(k i )

, where

T (m)

istheexpeted timespentonthe

point loationin theminimizationdiagramof

m

lines. Thus weobtainanex-

petedrunning-timeof

O(log 2 n)

intotal.

2

Weremarkthatsomespeialasesareleftoutinthisdisussionforbrevity

(e.g.,pointsthatare ontainedin

H

),buttheyareompletelyhandled inour

software.

5 Implementation Details

OurimplementationisbasedonCgal,whihfollowsthegeneri-programming

paradigm[3℄. Algorithmsareformulatedandimplementedsuh that theyare

abstratfromtheatualtypes,onstrutions,andprediates. Thus, theimple-

mentation of every algorithm and data struture in Cgal is parametrized by

a so-alledtraits lass [28℄, in whih these funtionalities are dened. Inpar-

tiular,auseranemployanalgorithmwithhisowntypes,onstrutions,and

prediatesby providinghis own traitslass. This wayitispossibleto ahieve

a great amount of exibility. At the extreme, it is possible to even partially

hange the nature of an algorithm, as we do here for the three-dimensional

lowerenvelopelass[32,Ÿ8.5℄.

The ore of our implementation is the traits lass for the lower envelope

algorithm, whih also needstobeavalidtraits lassforCGAL's arrangement

pakage. Therequiredfuntionalitiesbythearrangementpakageareprovided

bythetraits lasspresentedin [4℄. Theapproahredues allonstrutionand

prediates to ylindrial algebraideompositions of the plane for oneortwo

urves.Theapproahusesadditionalresultantomputationthatprojetsinter-

setion pointsontothe

u

-axis, that is,

u

-oordinatesof intersetionpointsare represented as real roots of this univariateresultant polynomial. Thereafter,

thebersabovetherootsareinvestigatedin orderto determinethearofthe

urveonwhih theintersetion takesplae.

Weessentiallywraptheabovetraitslassandadd therequiredfuntional-

itiesby theenvelopealgorithm;see also Setion 3. Inasewe detetthat the

urrentframe is notgeneri an exeption is thrown, whih is then aught by

ourprimary lassthat omputesanewframeandrestartstheomputation of

(13)

(a) (b) ()

Figure 2: Diagramsarelippedbyasphere foronveniene. (a)

VD

of

4

lines, obtained

byrotating onelinearound the

z

-axis. Allbisetorsmeetinthat axis. (b)

VD

of

4

lines

interseting inonepoint. ()

VD

of

4

lines,twolinesintersetandtheothersareparallelto

eahofthem,respetively.

the ell. Foreah Voronoi ellwe keep aseparate instane ofthe traits lass,

whihisusedforbothplanes. Thisallowsahingofrelevantresults.

Approximationofthethree-dimensionaloordinatesofavertex,isbasedon

multi-preision oating-point interval arithmeti (MPFI) [8, Ÿ8℄. Sine this is

aertied approximation,weobtainabounding boxthat ontainsthevertex.

This ouldbe usedto easily establishtheadjaeny among lowerdimensional

ells. Forinstane, let

v

denote a vertex in aminimization diagram

M

. The

labelof

v

pointstoallotherminimizationdiagramsthatontainarepresentation of it. Let

M

be oneof these diagrams and

v

be the representationthat we wish to ndtherein. We oulduseasimilar approah to theoneusedin [13℄:

Byusingthelabels,weidentifyallpossibleandidatesin

M

. Thissetontains

onlyupto

8

representationsandontainsatleast

v

. Weomputeprogressively more preise bounding boxes for allandidates until only one (the one of

v

)

overlapsthebounding boxof

v

.

Our implementation an handle arbitrary rational lines, in partiular, it

anhandle all possible degenerateases. Figure 2depits degenerateVoronoi

diagrams. Eah mesh was generated using CGAL's pakagefor labeledmesh

domains[30℄. Theorale,whihisrequiredbythemeshgeneration,waswritten

suhthatitonlyutilizes(andtherebytests)ourpointloationstruture. Lower

dimensional features were approximated using the approah disussed above.

In order to ahieve sharp edges the proteting balls tehnique introdued by

Bolthevaetal.[9℄wasapplied. medit[29℄wasusedforthenalvisualization.

Sineweaim to eventually inorporate our ode into aCgal pakagethe

software is developed within the revision ontrol system of the projet. All

experimentswithin thissetion wherearriedoutonan internalCgal release

CGAL-3.7-I-27,whihalreadyomprisesalltheneessaryalgebraitools[4℄.

However, the trapezoidal map is urrently not available for minimization di-

agrams due to ongoinghanges in the arrangementpakage (it is antiipated

soon),whih foresus toresortto asimplerpointloationstrategiesfornow.

Finally,wepresentpreliminaryresultsobtainedwithoursoftware. Thepoint

loationstrutureasitisdisussedinSetion4leavestheratio

k

amonglevels

(14)

N \k

2 4 8 16 20 24 28

+∞

16 6.48 4.30 4.34 N/A N/A N/A N/A 3.94

36 8.09 6.33 5.33 5.67 N/A N/A N/A 5.62

64 9.77 6.42 5.73 6.07 6.00 6.12 6.83 6.63

100 9.87 7.22 6.18 6.45 6.97 6.83 7.13 7.43

Table1: Averagenumberofvisitedellsperquery,where

k

denotestheratioofthehierarhy

and

N

the number oflines. Entriesfor

n < 2k

arerepesented bythe lastolumn. Tothe

rightisdepitedaVoronoidiagramof

5

parallellines.

undetermined. Inorder to show theimpat of

k

wereated random instane

of parallel lines with oeients in the range

[0, 2 10 ]

.5 For eah instane, we

reated

10

Voronoidiagramhierarhies,whihwherequeriedwith

1000

random

pointsin

[0, 2 10 ] 3

eah.

Table 1shows theaverage numberof visited ellsperquerydepending on

thenumberoflinesandthehosenvaluefor

k

. Thelastolumnshowsthepure

walkwithoutahierarhy,whih suggestsanaveragequery timein

O( √ n)

, as

onemayalso expet due to resultsin [12℄. For largerinstanes, itseemsthat

hoosing

k

equalto

8

isappropriate.

6 Conlusions

We havepresented anexat, omplete, and thus robust, algorithm that om-

putes the Voronoi diagram of arbitrary rational lines in

R 3

. The algorithm

requires

O(n 3+ε )

time and spae, where

n

is the number of lines. The in-

trodued datastruture permits to answerpoint loation queriesin

O(log 2 n)

expeted time. The implemented prototype is exat and an handle all de-

generateases. Wereferto http://ag.s.tau.a.il/projets/internal-

projets/3d-lines-vor/projet- page forthe mostreentversionand sup-

plemental material.

Thealgorithmisintentionallydesignedsuhthatitavoidstediousasedis-

tintions, whih makesit implementable,maintainable and, in partiular, ex-

tensibleto otherprimitivessuhaspoints,line segments,andtriangles. Thus,

weonsiderourapproahasamajormilestonetowardstheomputationofthe

Voronoidiagramofpolyhedrainthreedimensions.

Moreover,weexpetthatitwillpavethewaytodevisingathree-dimensional

variantofthevisibility-Voronoiomplex[34℄,astruturethatenablestotrade-

o learaneand pathlength in robot motionplanning,and hasproved tobe

espeially usefulin theplane.

Ourapproahmayalsobegeneralizedtospheres(seealso[20℄)whihwould

open the doorfor innovative solutionsto entral problems in StruturalBiol-

ogy[24,35℄.

Aknowledgments: Theauthors thankS. Lazard andM. Yvineforfruitful

disussionsandMoniqueTeillaudforhertranslationof theabstratto Frenh.

5

Sinethetrapezoidalmapisnotyetavailableforenvelopes,wehadtoresorttoinstanes

thatkeeptheomplexityofaellsmall.

(15)

Referenes

[1℄ P. K.Agarwal, O.Shwarzkopf,andM. Sharir. The overlay oflowerenvelopes

anditsappliations. Dis.Comput. Geom.,15(1):113,1996.

[2℄ F. Aurenhammerand R.Klein. Voronoi diagrams. InJ.Sak and G.Urrutia,

editors,Handb.Comput. Geom.,hapter5,pages201290.Elsevier,2000.

[3℄ M.H.Austern. Generi ProgrammingandtheSTL. Addison-Wesley,1999.

[4℄ E.Berberih,M.Hemmer,andM.Kerber. AnalgebraikernelforCGAL. Sub-

mittedtoESA,2010.

[5℄ E. Berberih,M. Hemmer,L. Kettner,E.Shömer,and N.Wolpert. Anexat,

ompleteandeientimplementationforomputingplanarmapsofquadriin-

tersetionurves.InPro.21stAnnu.ACMSymp.Comput.Geom.,pages99106,

ACMPress, 2005.

[6℄ H. Blum. A transformation for extratingnewdesriptors of shape. InModels

forthePereption ofSpeeh andVisualForm,MITPress, 1967.

[7℄ J.-D.BoissonnatandC.Delage. ConvexhullandVoronoidiagramofadditively

weightedpoints. InESA,pages367378,2005.

[8℄ J.-D.BoissonnatandM.Teillaud,editors. EetiveComputationalGeometryfor

Curves andSurfaes. Springer-Verlag,MathematisandVisualization,2006.

[9℄ D. Boltheva, M. Yvine, and J.-D. Boissonnat. Feature preserving Delaunay

mesh generation from 3D multi-material images. Computer Graphis Forum,

2009.

[10℄ T.Culver,J.Keyser,andD.Manoha. Exatomputationofthemedialaxis of

apolyhedron. Computer AidedGeometriDesign, 21(1):6598,2004.

[11℄ O.Devillers. ImprovedinrementalrandomizedDelaunaytriangulation.InPro.

14thAnnu.ACMSymp.Comput.Geom.,pages106115,ACMPress,1998.

[12℄ L.Devroye,C.Lemaire,andJ.-M.Moreau. Expetedtimeanalysisfordelaunay

pointloation. Computational Geometry,29(2):6189, 2004.

[13℄ L. Dupont, M. Hemmer, S. Petitjean, and E. Shömer. Complete, exat and

eient implementationfor omputing the adjaeny graphof anarrangement

ofquadris. InPro.15thAnnu. Eur.Symp.Alg.,volume4698ofLNCS, pages

633644,Springer,2007.

[14℄ H.EdelsbrunnerandR.Seidel.Voronoidiagramsandarrangements.Dis.Com-

put.Geom.,1:2544,1986.

[15℄ I. Z. Emiris and M. I. Karavelas. The prediates of the Apollonius diagram:

Algorithmianalysis and implementation. Comput. Geom.Theory Appl., 33(1-

2):1857,2006.

[16℄ I.Z.Emiris,E.P.Tsigaridas,andG.M.Tzoumas.TheprediatesfortheVoronoi

diagramofellipses.InPro.22ndAnnu.ACMSymp.Comput.Geom.,pages227

236,ACMPress,2006.

[17℄ H.Everett,C.Gillot,D.Lazard,S.Lazard,andM.Pouget.TheVoronoidiagram

ofthreearbitrarylinesin

R 3

.InAbstratsof25thEur.WorkshopComput.Geom.,

2009.

[18℄ H. Everett, S.Lazard, D. Lazard, and M. S.E. Din. The Voronoi diagram of

three lines. InPro. 23rd Annu. ACM Symp. Comput. Geom., pages 255264,

ACMPress, 2007.

[19℄ D.Halperin,L.E.Kavraki,andJ.-C.Latombe. Robotis.InJ.E.Goodmanand

J.O'Rourke,editors,Handb.Dis.Comput.Geom.,hapter48,pages10651093.

Chapman&Hall/CRC,2ndedition,2004.

(16)

[20℄ I. Hanniel and G. Elber. Computing the Voronoi ells of planes, spheres and

ylindersin

R 3

. Comput. AidedGeom.Des., 26(6):695710,2009.

[21℄ I. Haran and D. Halperin. An experimental study of point loation inplanar

arrangementsinCGAL. ACMJournalofExperimentalAlgorithmis,13, 2008.

[22℄ M. I. Karavelas. A robustand eent implementation for thesegment Voronoi

diagram. In Int. Symp. on Voronoi Diagrams in Si. and Engineering, pages

5162,2004.

[23℄ M. I.KaravelasandM.Yvine. DynamiadditivelyweightedVoronoidiagrams

in2D. InPro.10thAnnu.Eur.Symp.Alg.,pages586598,Springer,2002.

[24℄ D.-S.Kim,J.Seo,D.Kim,Y.Cho,andJ.Ryu.Thebeta-shapeandbeta-omplex

for analysis of moleular strutures. In M. L. Gavrilova, editor, Generalized

Voronoi Diagram: A Geometry-Based Approah to Computational Intelligene,

volume158ofStudiesinComputationalIntelligene,pages4766.Springer,2008.

[25℄ V.KoltunandM.Sharir.3-dimensionalEulideanVoronoidiagramsoflineswith

axednumberoforientations. SIAMJ.onComputing,32(3):616642,2003.

[26℄ V.Milenkovi. RobustonstrutionoftheVoronoidiagramofapolyhedron. In

Pro.5thCanad.Conf.Comput. Geom.,pages473478,1993.

[27℄ K.Mulmuley. A fast planarpartition algorithm, I. InPro.29th Annu. IEEE

Sympos.Found. Comput.Si.,pages580589,1988.

[28℄ N.Myers. Traits: Anewandusefultemplatetehnique. C++Gems,17,1995.

[29℄ Pasal J.Frey. MEDIT:AninterativeMeshvisualizationSoftware. Tehnial

ReportRT-0253,INRIA,De2001.

[30℄ L.RineauandM.Yvine.3Dsurfaemeshgeneration.InCGALEditorialBoard,

editor,CGALUserandReferene Manual.3.5 edition,2009.

[31℄ M.Sharir. Almosttightupperboundsforlowerenvelopesinhigherdimensions.

Dis.Comput. Geom.,12(1):327345,1994.

[32℄ TheCGALProjet. CGALUserandRefereneManual.CGALEditorialBoard,

3.6 edition,2010. http//www.gal.org/.

[33℄ J.vonzurGathenandJ.Gerhard. ModernComputer Algebra. CambridgeUni-

versityPress,1999.

[34℄ R.Wein,J.P.vandenBerg,andD.Halperin.Thevisibility-Voronoiomplexand

itsappliations.ComputationalGeometry: TheoryandAppliations,36(1):6687,

2007.

[35℄ E. YaeandD. Halperin. Approximating thepathway axisand thepersistene

diagram of a olletion of balls in 3-spae. In Pro. 24th Annu. ACM Symp.

Comput. Geom.,pages260269,ACMPress,2008.

[36℄ C.-K. Yapand T. Dubé. Theexat omputation paradigm. In D.-Z. Du and

F. K. Hwang, editors, Computing in Eulidean Geometry, volume 1 of LNCS,

pages452492.WorldSienti,2ndedition,1995.

(17)

A Projetion of Bisetors and Trisetors

Reall, from thedisussion in Setion 3that the parametrization

X (u, v, r)

is

dened with respetto a loal frame

F = { − → b 1 , − → b 2 , − → b 3 }

and that the plain

H

,

whihisorthogonalto

→ b 3

,isnotontainedin

X

.

Ingeneral this is not aproblem sinea urve (trisetor)that transversely

intersets

H

appearsasasimplevertialasymptoteinthe

uv

-parameterspae.

Forinstane,ifaprojetedurveleavesoneplaneasavertialasymptoteatplus

innity,thenitreappearsatminusinnityontheotherplane. Ifthisistheonly

asethathappensitislearthatthereisaone-to-onorrespondeneamongthe

vertialasymptote (andthus edges) in thetwo minimization diagrams,whih

makesitpossibletogluethemtogether.

Weallaframeforwhihthisispossibleageneriframe,morepreisely:

Denition5 (GeneriFrame) Wedenetheframe

F

tobe generiif(i)no

1-dimensional omponents of the trisetor (a oni or a line) 6

are ontained

in

H

,(ii)everyintersetionofatrisetorwith

H

isatransversalintersetion, (iii)notwourves(trisetors)intersetin

H

,and(iv)the intersetionsof

H

with allbisetors

B 0i

areregular.

Inthissetion weanalyze theasesthat weenounterwhile projetingbi-

setorboundariesandtrisetorurves. Inpartiular,wedisusshowwedetet

anon-generiframe.

A.1 Projetion of Bisetors Boundary

TheaseshereorrespondtotheasedistintioninProposition1.

Generi Case Let

B 0i ∈ Q[x 1 , x 2 , x 3 ]

be the polynomial representing the bisetorbetween

ℓ 0

(thebaseline)andsomeotherline

ℓ i

. Inthegeneriase

B 0i

representsahyperboliparaboloidthat isdenein almostalldiretions ofthe

projetion. Theonlyexeptionareraysthat areperpendiularto

ℓ i

andpoint

awayfrom

ℓ i

. Thesediretionsarerepresentedbyahorizontal(representingrays in thesamediretion)line. Thelineis haraterizedbytheleadingoeient

of

B 0i ( X (u, v, r))

with respetto

r

. Oneah ofthe

uv

-planeweinterpret this

surfae as one or two

uv

-monotone surfaes. On the plane that ontains the

above line, we split the bisetor into two

uv

-monotone surfaes; on the other

planethereisexatlyone

uv

-monotonesurfae.

Parallel Lines Inase

ℓ 0

and

ℓ i

areparallel,

B 0i

isof degree

1

sineitrep-

resents aplane. Theleading oeientof

B 0i ( X (u, v, r))

with respetto

r

is

a horizontal line that represents all rays that do not interset

B 0i

. That is,

oneah ofthe

uv

-planesweobtainone

uv

-monotonesurfae,whose projeted

boundaryisthisline. Theproperhalfspaeanbedeterminedbyasimpleray

shoot.

6

Notethataubiannotbeontainedin

H

.

(18)

IntersetingLines Inase

ℓ 0

and

ℓ i

interset,thebisetor

B 0i

degenerates to twoplanesthat intersetalong asingularrational line

ℓ s

, whih is perpen-

diular to

ℓ 0

and

ℓ i

. Let

(u 0 , v 0 ) ∈ Q 2

be theparameter valuesfor that line.

All lines

X (u 0 , v, r), v 6 = v 0

have a double intersetion with

B 0i

. The lines

X (u, v 0 , r), u 6 = u 0

donot interset

B 0i

, sine theyare parallel to it. Thepa-

rameterspaeissplitupalong

v = v 0

and

u = u 0

,whihresultsin

uv

-monotone

surfaes,

4

onethepositiveand

4

onthenegativeside.

Ensuringa GeneriFrame Forallasesitholdsthattheonstrutionmay

trigger arestart of theomputation ifone of the horizontal lines is not seen,

thatis,iftheleadingoeientof

B 0i ( X (u, v, r))

withrespetto

r

hasdegree

0

.

A.2 Projetion of Trisetors

Besides the exatness of the method desribed in Setion 3.1, it has the ad-

vantage that it is general and forgoesa huge ase distintion. In partiular,

wedonotfatorize thepolynomialinto itsfators thatrepresentthe dierent

omponentsmentionedinTheorem 2.

Twobisetorsareomparedabove(below)a

u

-monotonear

c u

,

7

onthepos-

itive(negative) planeasfollows. Forasuientlylose rationalpoint

p(u 0 , v 0 )

above(below)

c u

, the orrespondingline is substituted into thetwobisetors.

Thisisatmosttwoquadratipolynomialsin

r

,eahhavingatmostonepositive

and one negativeroot. If present, the positive (negative) roots from the two

bisetors are ompared. In ase that no root is present, there is no bisetor

above

p

andtheomparisonisnotrequired.8

Aspeial treatmentisgivenfortheaseofthreeoplanaronurrentlines.

Thetrisetorinthisaseisasingleperpendiularlineto

ℓ 0

. Hene,itsproje-

tionisjustarationalpointthatisvalidforbothsides.

Ensuring a GeneriFrame First,itishekedthatno1-dimensionalom-

ponentisontainedin

H

,inthatase

degree(res)

isneessarilylessthan

2d 1 d 2

but the degreemay alsodrop in fewother ases. Thus ifthedegree issuspi-

ious, we simply interset

B 1

and

B 2

with

H

. There is no 1-dimensional omponentif the degreeof the

gcd

of thetworesulting bivariate polynomials is onstant. Moreover, we hek that the projetion has only simple vertial

asymptotebyhekingthattheleadingoeientofthesquarefreepartof

res

issquarefree. Thisensuresthatthetrisetorintersets

H

transverselyalways.

7

Aboveistheareatotheleftoftheurvewhengoingfromitslexiographiallysmaller

toitslexiographiallylargerend.

8

Thishappensinasethebisetorisjustasimpleplane.

(19)

Contents

1 Introdution 3

2 Preliminaries 4

2.1 PropertiesofBisetorsandTrisetors . . . 4

2.2 LowerEnvelopeAlgorithm. . . 5

3 Computinga Voronoi Cell 6

3.1 ParametrizationandProjetion . . . 6

3.2 LowerEnvelopePrediates . . . 7

3.3 Complexity . . . 8

4 Fast Point Loation 8

5 ImplementationDetails 9

6 Conlusions 11

A Projetion of Bisetorsand Trisetors 14

A.1 ProjetionofBisetorsBoundary . . . 14

A.2 ProjetionofTrisetors . . . 15

(20)

2004, route des Lucioles - BP 93 - 06902 Sophia Antipolis Cedex (France)

Centre de recherche INRIA Bordeaux – Sud Ouest : Domaine Universitaire - 351, cours de la Libération - 33405 Talence Cedex Centre de recherche INRIA Grenoble – Rhône-Alpes : 655, avenue de l’Europe - 38334 Montbonnot Saint-Ismier Centre de recherche INRIA Lille – Nord Europe : Parc Scientifique de la Haute Borne - 40, avenue Halley - 59650 Villeneuve d’Ascq

Centre de recherche INRIA Nancy – Grand Est : LORIA, Technopôle de Nancy-Brabois - Campus scientifique 615, rue du Jardin Botanique - BP 101 - 54602 Villers-lès-Nancy Cedex

Centre de recherche INRIA Paris – Rocquencourt : Domaine de Voluceau - Rocquencourt - BP 105 - 78153 Le Chesnay Cedex Centre de recherche INRIA Rennes – Bretagne Atlantique : IRISA, Campus universitaire de Beaulieu - 35042 Rennes Cedex Centre de recherche INRIA Saclay – Île-de-France : Parc Orsay Université - ZAC des Vignes : 4, rue Jacques Monod - 91893 Orsay Cedex

Éditeur

INRIA - Domaine de Voluceau - Rocquencourt, BP 105 - 78153 Le Chesnay Cedex (France)

http://www.inria.fr

Références

Documents relatifs

Since the Voronoi diagram of smooth convex pseudo-circles is an abstract Voronoi diagram, given 3 sites, there may exist at most one CCW Voronoi disk and at most one CW Voronoi

Remark 3.5. To determine whether the sailboat follows the desired di- rection, there could be many criteria using the sensor data about speed, direction or position of the boat.

Any set of points, known as Voronoi sites, in a given n-dimensional Euclidean space E defines a Voronoi diagram, i.e., a partition of that space into Voronoi cells: the

The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.. L’archive ouverte pluridisciplinaire HAL, est

The insertion of a new object A in the current Voronoi diagram V(A) involves the following steps: (1) find a first conflict between an edge of V (A) and A or detect that A is hidden

Remark 3.5. To determine whether the sailboat follows the desired di- rection, there could be many criteria using the sensor data about speed, direction or position of the boat.

Body text region is then extracted from stele images by the selec- tion of appropriate adjacent Voronoi regions based on the information about the thickness of neighboring

For each vertical text line extracted from the body text region using vertical projection profile, area Voronoi diagram is employed to represent the neighborhood of connected