SEGMENTATION AND MODELLING OF FULL HUMAN BODY

SHAPE FROM 3D SCAN DATA: A SURVEY

Naoufel Werghi

Dubai University College

Dubai, UAE

Keywords:

Whole Human Body Scanner, Human Body Scan Segmentation, Human Body Shape Modelling.

Abstract:

The recent advances in full human body imaging technology illustrated by the 3D human body scanner (HBS),

a device delivering full human body shape data, opened up large perspectives for the deployment of this

technology in various ﬁelds (e.g. clothing industry, anthropology, entertainment). Yet this advance brought

challenges on how to process and interpret the data delivered by the HBS in order to bridge the gap between

this technology and potential applications. This paper surveys the literature on methods, for human body scan

data segmentation and modelling, that attempted to overcome these challenges. It also discusses and evaluated

the different approaches with respect to several requirements.

1 INTRODUCTION

The last decade has witnessed the emergence of new

3D imaging devices capable of capturing the entire

shape as well as the appearance of the human body

(HB). A human body scanner (HBS) is a device that

generates a three-dimensional ”point cloud” from the

subject’s frame, i.e. a constellation of 100,000 -

200,000 points generated by the body’s surface. This

data is saved into a simple digital format and can eas-

ily be converted to the most common computer-aided

design formats. This development opens up new per-

spectives for human body scanners in diverse ﬁelds.

In anthropometric surveys which require collect-

ing more than one hundred body measurements from

a large population (thousands of individuals), man-

ual measurement is time-consuming, error-prone, and

very costly. In contrast, HBS technology drastically

reduces the cost and duration of the surveys. Indeed,

it permits rapid capture of the body shape - seconds

versus minutes for a measurement by hand - and of-

fers more consistent measurements. In addition to

the speed of data collection, HBS offers reusability

of data, as the scan data actually replaces the scanned

subject. In effect, once extracted, the scan data can

be used again and again to gather additional informa-

tion, whereas a subject measured once by traditional

means is no longer available for future reference.

Clothing design and human engineering are also

potential beneﬁciaries of HBS scan technology (Pa-

quette, 1996). The clothing industry is presently tar-

geting custom apparel design, commonly referred to

as ”apparel on demand” which aims to produce cloth-

ing designed and ﬁtted to an individual’s size and pro-

portions. This will permit better ﬁtting garments, par-

ticularly for individuals outside the normal size range,

thus reducing the cost of labour involved, and en-

suring a rapid response by substantially reducing the

time between measurement and delivery. HBS has

a positive impact on human engineering which re-

lies extensively on anthropometric databases (Caesar

project, Online), and which is involved in custom-

ﬁtting items to human surfaces. Such items are gener-

ally used by great numbers of people and include pro-

tective equipment such as helmets, seat belts, desks,

airplane and car seats.

HBS also opens up new applications in medicine

and health. Gyms can use HBS data to track the ef-

fects of diet and exercise regimens. Human shape data

bases will be useful for screening and survey tasks -

for instance, monitoring public health problems such

as obesity, and assessing child growth. HBS would be

of signiﬁcant interest particularly when use of other

standard medical tools like X-rays are precluded for

safety reasons. In fact, the relative low cost and non-

invasive nature of HBS, make it a promising potential

complement to current medical imaging technologies

used to assist medical diagnosis, such as Computed

Tomography Imagery (CTI) and Ultrasound Imagery

(USI).

189

Werghi N. (2006).

SEGMENTATION AND MODELLING OF FULL HUMAN BODY SHAPE FROM 3D SCAN DATA: A SURVEY.

In Proceedings of the First International Conference on Computer Vision Theory and Applications, pages 189-197

DOI: 10.5220/0001367401890197

 SciTePress

Further, there appears to be a deﬁnite future for

HBS in ﬁelds that involve VR applications, such as

movies, television production and games (in which

human-like characters appear as virtual actors, or

even interact with users, usually by taking the role of

an opponent). In such applications, HBS scan data

can be used to provide realistic and customized mod-

els.

It is to be noted that the technology of full human

shape capture is continually improving at the same

time as its cost is going down. This is taking place

so rapidly that novel and even unforeseen customized

application areas will be realizable in the near future,

considering the proliferation of personal computers

and different types of communication networks. For

example, a body scanner could be coupled to a PC and

the captured human body shape information could

then be used for any of a number of purposes, e.g.

electronic commerce, ordering ﬁtted items, or remote

medical diagnosis.

Such huge potential for the exploitation of HBS

raises important questions on how to process and in-

terpret raw captured human shape data in order to ex-

tract needed information and facilitate further use of

such data in various applications. In effect, the 3D

point cloud data produced by HBS itself is nothing

more than a collection of coordinate values, and no se-

mantic interpretation of this data is possible. In cloth-

ing design applications and anthropological surveys,

questions arise as to how to locate body landmarks

used as reference points. In medical applications, how

can the whole scan data set be partitioned into subsets

corresponding to the principal body parts? Or, more

globally, how will it be possible to effect an accurate

segmentation that preserves the topological and mor-

phological characteristics of the human body shape?

In applications involving modelling human body

shape, more pertinent questions have been raised over

the best way to represent the human body. In ef-

fect, while human body shape cannot be represented

by simple parameterised surfaces, such as planes and

quadrics, it is not a randomly-shaped free form sur-

face either, and thus practitioners are in search of a

representation that ensures maximum embodiment of

human body forms while at the same time exhibiting

an optimal trade-off between conciseness and expres-

siveness. As an example, will it in fact be possible to

deﬁne a typical surface shape for the arm which can

represent the full spectrum of shapes the human arm

takes? And what criteria should be used in the deﬁni-

tion of such representations?

It is hoped that this paper will shed some light

on the steps taken toward answering certain of these

questions. The next section offers a brief descrip-

tion of the technology employed in HBS. Section 3

contains a detailed overview of research done so far

on human body shape analysis and modelling, while

discussing and evaluating the different approaches in

light of the problems and questions mentioned above.

The paper concludes with some general observations

and potential research orientations.

2 HUMAN BODY SCANNER

TECHNOLOGY

Basically there are two categories of technology em-

ployed in HBS, laser-based technology and moir

fringing technology (Figure 1). Both are optical and

involve no direct contact. In laser-based technology,

which was developed by Cyberware (Cyberware, On-

line), a laser beam is projected from eight laser diodes

onto the body, scanning it from top to bottom. The

laser stripe, deformed by the body surface, is captured

by different cameras around the body and recorded in

a digital format. The captured data is basically the

location of the laser stripe with respect to the camera

reference. Afterwards a software program combines

separate data from each camera, using a so-called

triangulation technique, to produce a set of 3D data

points representing the body surface. The duration of

the scan is around 17 seconds.

Moir

e fringing technology employs a moir

e-based

light-projection system, known also as Phase Measur-

ing Proﬁlometry. This system was developed by the

textile and clothing technology corporation (TC

), in

Cary, North Carolina (TC2, Online). In this system, a

white light source is used to project contour patterns

(sinusoidal fringes) on the body surface. The con-

tour patterns distorted by curves in the body surface

are detected by a set of cameras arranged around the

subject and linked to a computer. The superimposed

deformed patterns thus generated interact with other

patterns used as reference points. These in turn form

fringes that describe the body surface contours. Sub-

sequently the data obtained from the separate fringes

are combined into a single reference yielding a cloud

in which the 3D data points represent the body sur-

face. Scan time in this technology is around two sec-

onds.

3 LITERATURE REVIEW

Approaches dealing with HBS scan data can be clas-

siﬁed into three themes, namely, human body land-

mark detection, HB scan data segmentation and hu-

man body shape modelling. However we mention that

some approaches in the literature touched more than

one theme. The next sections will discuss these three

themes in details.

VISAPP 2006 - IMAGE ANALYSIS

190

Figure 1: Basically a human body scanner is composed

of an arrays of cameras and projectors arranged in a square

or triangular fashion (a). The projectors produce patterns

on the body surface, which are captured by the cameras

and processed to generate data on the subjects shape and

skin colour. The cameras and the projectors operate syn-

chronously. Cyberware scanner (Cyberware, Online) (b)

employs laser beam projectors, whereas TC

-like scanners

(TC2, Online) use Moir

e fringes (c) (courtesy of Inspec:

www.inspec.com). (d): Cloud of 3D Data points represent-

ing the body surface. (e): A solid model of the body sur-

face. (f): The solid model mapped with real texture which

is provided by the scanner.

3.1 Human Body Landmark

Detection

Motivated by the need for automatic and accurate ex-

traction of body measurements in apparel design ap-

plications, the very ﬁrst approaches used special pat-

terns to mark body landmarks so that they could be

easily detected on images provided by the HB scan-

ner. The work of Geisen et al (Geisen, 1995) is an

illustrative example, in which adhesive patterns are

stuck to the anatomical landmarks of the body. The

positions of these patterns are detected in the image

delivered by the cameras and then mapped to the 3D

data points to obtain their 3D positions. However,

one drawback is that the identiﬁcation of the detected

landmarks relies on prior knowledge of the scanned

portion of the body and its relative position. Be-

sides, using these patterns is a complication in the pa-

tient scanning assessment. Therefore, these kinds of

approaches were quickly abandoned. Some authors

turned towards manual solutions, e.g. Pargas et al

(Pargas, 1997), who developed a software package in

which sliced body scan data is edited and the body

measurements are then extracted manually. Attempts

made to automate measurement extraction were based

on rough approximations of the position of the rel-

evant body landmarks; Therefore margin of error in

these measurements was signiﬁcant. Using the same

application, Jones et al (Jones, 1995) focused on the

torso. A set of cross-sectional slices is manually se-

lected in the vicinity of the key anatomical landmarks

in the torso. These slices are ﬁtted together and then

interpolated to generate a NURBS approximation (a

kind of CAD format). This technique ensures a vi-

able trade-off between compactness and surface de-

tail preservation. The approach remains particularly

suited to clothing design however, and the process

needs a great deal of manual intervention, although

some attempts were made to automatically separate

the torso and the upper part of the arms (Li, 1997).

3.2 Human Body Scan Segmentation

Automatic segmentation of the human body into its

functional parts was ﬁrst studied by Nurre (Nurre,

1997; Nurre, 2000). In his pioneering work, he ap-

proximated the body structure by a six-stick template

representing the head, the two arms, the two legs and

the torso. The goal was to segment the body into

six segments corresponding to these parts. This ap-

proach combines global shape description, in partic-

ular, moment analysis and local criteria of proximity,

which are derived from prior knowledge of the rel-

ative position of the body parts in standard posture

(standing body with arms held at the sides). The scan

data is organized into slices of data points. These

horizontal slices are stacked vertically and the data

points are assigned to different body parts accord-

ing to the topology of the slices and their position on

the body Figure (Figure 2). While this work made

considerable headway towards the automatic decom-

position of HB scan data, it has been criticized for

imposing the requirement of limiting body poses to

strict standard postures and for its lack of robustness

against noise, gaps in the data, and variation of shape

and posture of the HB. There have been many sub-

sequent attempts to improve’s Nurre’s approach, with

several efforts to enhance the localisation of key land-

marks of the human body. For example, Decker et al

(Dekker, 1998) improved the localisation of the key

landmarks of the HB by applying differential opera-

tions on slice shape attributes, with respect to body

shape modelling. They also proposed B-spline ap-

proximation of the torso (Douros,1999). Although a

degree of improvement resulted from this work, this

related approach could not remedy the limitations of

Nurre’s approach. Wang et al ( Wang, 2003) proposed

a new approach based on a framework employing

fuzzy logic. Their segmentation technique involved

local curvature analysis of the slices and operated on

mesh data that must of necessity undergo several pre-

processing stages. However, this again was restricted

to standing postures. The overall performance of this

approach remains identical to that of Nurre’s. Despite

the obvious improvements illustrated by these previ-

SEGMENTATION AND MODELLING OF FULL HUMAN BODY SHAPE FROM 3D SCAN DATA: A SURVEY

191

Figure 2: Segmentation of the HB data in the case of a

standard standing posture (a). The scan data is sliced hor-

izontally (b). Afterwards the slices are analyzed topologi-

cally inferring the knowledge of the human body template

(c), for example, a slice having two separated closed curves

must represent data points generated by the legs, a slice con-

sisting of three closed curves must belong to the torso/arms

area and a slice with two joined closed curves is assumed to

correspond to the transition between the legs and the torso

(at the level of groin). The output of this stage is a seg-

mented scan data (d).

ous works, none could fully meet the requirements

that one would desire in a segmentation approach. In

effect, to be of practical utility, HB scan data segmen-

tation must: 1) be robust to variation in the body sur-

face shape stemming from biological factors such as

age, genetics, etc; 2) be able to cope with changes

of body posture, bearing in mind that a full recovery

of the human body requires more than one posture

(Brunsman, 1997); and 3) cope as well with diversity

of the scan data sources as well as data deﬁciencies

and corruption.

To solve the issue of the stability of the HBS scan

data segmentation with respect to posture changes,

some authors proposed the recovery of body pos-

ture from the scan data. Then once the posture is

identiﬁed, the underlying information on the loca-

tion of body parts can be exploited in the segmen-

tation. Along with this beneﬁt, posture recognition

is also of interest for other applications such as hu-

man motion and gesture analysis (Lin, 1999), where

the knowledge of posture intervenes in the initialisa-

tion of the tracking algorithm, and the management of

three-dimensional and anthropometric databases (Pa-

quet, 2000), where posture recognition is vital for re-

trieval and classiﬁcation tasks (Paquet, 2001).

Werghi et al (Werghi, 2002; Werghi(1), 2005) con-

cretised the idea of exploiting the knowledge of pos-

ture in a two-phase algorithm. The ﬁrst phase in-

volves a Bayesean classiﬁcation approach, employ-

ing wavelet coefﬁcients-based descriptors, by which

the posture is identiﬁed. In the second phase a scan

data slicing and analysis embodying the knowledge of

posture is applied. This approach had the advantage

of breaking the barrier of strict standard postures. But

despite this advance, it has been abandoned because

the spectrum of postures covered by this approach

Figure 3: Segmentation of HB scan data for an arbitrary

posture (a). Computation of the Level-sets (sets of data

points located at the same geodesic distance with respect

to a source point)(b). Construction of the Discrete-Reeb

Graph (c) (a graph that encompass the human body tem-

plate, and which is invariant to posture changes). The seg-

ments and the joints of the graph are then mapped to the

scan data to extract the different body parts (d).

was limited, since as a general principle, a recogni-

tion technique, whatever its power, can only deal with

a ﬁnite number of postures, and also because it could

not cope effectively with scan data corruption.

In response to these challenges Xiao et al (Xiao(1),

2003; Xiao(2), 2003) proposed a robust computa-

tional topology framework that copes well against

scan data corruption and diversity of the scan data

source. Basically, this framework permits building

a skeleton-based representation (known as the Reeb-

Graph in the topological community) that encodes the

human body template as well as critical points rep-

resenting key body landmarks, such as the armpits

and the torso. What was new in this technique was

the extension of the Reeb-Graph concept to the dis-

crete space. This new version, dubbed the Discrete

Reeb-Graph, and deﬁned on the basis of the connec-

tivity between discrete points and curves, can oper-

ate directly on point cloud data without any special

pre-processing. In the earlier version, the approach

could only cope with moderate variations on standard

posture, however this limitation was overcome in a

subsequent version (Xiao, 2004; Werghi(2), 2005) by

employing the geodesic distance (the closed distance

between two points on a surface) to construct the Dis-

crete Reeb-Graph. Being invariable to rigid transfor-

mations and isometric deformations, the geodesic dis-

tance implies a DRG construction impervious to hu-

man body movements and thus allowing a stable seg-

mentation with respect to posture changes (Figure 3).

This framework fulﬁls the three requirements cited

above, and has in fact proven to be able to segment

real-world human scans in arbitrary postures without

referring to any detailed human heuristics. Further, it

exhibits robustness against scan data deﬁciencies and

diversity of scan sources. The output of this process

consists of 5 sets of data points. each corresponding to

a major body part, i.e. the torso (including the head),

arms and legs.

VISAPP 2006 - IMAGE ANALYSIS

192

Figure 4: (a) superqadric-based model of a human body-

like shape doll [24]. (b) Example of a metaball modelling,

using a sphere primitive [34].

3.3 Human Body Shape Modelling

Historically the modelling of human body-like shapes

(e.g. dolls, mannequins) dates from the pioneering

work of Marr et al (Marr, 1976), who developed

a hierarchical generalized-cylinder representation in

which each part of the body is represented by a hi-

erarchically decomposable set of cylinder-like prim-

itives connected at their ends into a ﬂeshed-out stick

ﬁgure. Continuing in the same scope of application,

Pentland (Pentland, 1990) went further in complexity

by deﬁning a humanoid shape by a conjunction of su-

perquadric volumes, each associated with a body part.

Motivated by shape description rather than recogni-

tion, Terzopoulos et al (Terzopoulos, 1991) also used

superquadric primitives but, instead of a parametric

representation as for Pentland, they adopted a mesh-

surface model (Figure 4.a). In these two last studies,

all the data is ﬁtted to a set of generic superquadric

primitives and no surface shape analysis for detect-

ing the different body parts was involved. This prob-

lem was tackled by other authors (Borges, 1993;

Dion, 1997; Ferrie, 1993; Trucco, 1991) for whom

the measurement data is decomposed into sets corre-

sponding to the different segments of the humanoid

body. These sets are ﬁtted afterward either to quadrics

(Trucco, 1991), generalized cylinders (Dion, 1997)

or to superquadrics (Borges, 1993; Ferrie, 1993). In

the last two studies, the points of discontinuity in

the range data are ﬁrst detected, then dynamically

grouped into contours using an energy-minimization

process of deformable curves (snakes (Kass, 1988)).

The contours thus obtained deﬁne the separations be-

tween areas associated with the body parts. Certain

other papers attempted to address problems related to

articulated structure of the human shape, more specif-

ically registration of data corresponding to different

body postures (Ashbrook, 1999).

Without a doubt, this considerable body of re-

search has advanced the state of the art of modelling

articulated objects, and the theories developed are

a workable framework for approaching real human

body shapes. However, tackling the case of real hu-

man body scan date appears even more challenging,

ﬁrstly because the body shape is both articulated and

malleable (as opposed to the rigid human-like shape

treated in the above works) and secondly because the

scan data is by nature non-uniformly sampled and

may exhibit gaps and noise corruption. It has there-

fore been necessary to explore new techniques in or-

der to formulate approaches that are better able to

cope with these challenges.

Modelling of HBS data can be divided into two

types: static modelling, wherein the body is not sup-

posed to move, but local surface deformation is al-

lowed -body scans examined in this category corre-

spond generally to the standard static posture-; and,

secondly, dynamic body modelling, wherein the aim

is to model changes in shape as the body moves. The

next two sections of this paper will describe these two

approaches.

3.3.1 Static Body Modelling

Certain authors in this area were inspired by implicit

surface models developed for human-like shapes, for

instance (Shen, 1995; Matsuda,1999) used different

variants of a metaballs concept, deﬁned as iso-surface

(equi-potential surface) of a ﬁeld function. This con-

cept was ﬁrst introduced by Muraki (Muraki, 1991)

to model anatomic surfaces from 3D range data. Ba-

sically this type of modelling consists of ﬁtting group

weighted layered metaballs to the scan data within an

optimisation process. The metaball is quadric prim-

itve, either a sphere or an ellipsoid (Figure 4.b). In

this process both the weights and the number of meta-

balls determine the precision of the model. Implicit

surfaces are however difﬁcult to model and animate

interactively because of their considerable require-

ments in terms of calculations and human interven-

tions. Moreover, they cannot handle human shape de-

tails accurately, at least not without excessive compu-

tational cost. Besides, these methods fall victim to the

classic optimisation issues, such as initialisation and

the local minima.

Fuelled by these concerns, researchers instigated

another alternative called the conformation-based ap-

proaches . The principle of such approaches consisted

in coupling two models, namely, a template model

that encapsulates the coarse shape, and a detail model

that encompasses the local surface deformations.

Conformation-based approaches came in two ver-

sions, 2D-3D ﬁtting (Hilton, 1999; Starck, 2001;

Kakadiaris,1998) and 3D-3D ﬁtting (Ju, 2000; Ju,

2001)(Figure 5). In the ﬁrst version, 3D human body

model is deformed to ﬁt 2D body silhouettes extracted

from a set of views; whereas in the second version, a

generic 3D mesh model is ﬁtted to 3D scan data via a

a two-phase algorithm that consists of global and lo-

SEGMENTATION AND MODELLING OF FULL HUMAN BODY SHAPE FROM 3D SCAN DATA: A SURVEY

193

Figure 5: 3D-3D conformation technique: the 3D model

(b) is ﬁtted to the scan data (a) , the result (c) is a human

body model that embeds the shape of the scanned human

body. (d,e and f) the same technique applied for the case of

a human head.

cal mapping. Global mapping involves a reﬁned ver-

sion of segmentation technique (Nurre, 1997), which

is used to identify the different body segments (up-

per and lower arms, upper and lower legs and torso)

and hence establishing correspondences between the

scan data parts and the generic model parts. This

method inherited however the drawback of (Nurre,

1997), namely the restriction to standard posture. Lo-

cal mapping employs a closest-point correspondence

technique inspired from (Ashbrook, 1998).

In the same vein, other authors (Allen,2003; Seo,

2003) developed a more elaborated approach whereby

the global shape is deﬁned explicitly by human skele-

ton structure. The ﬁrst phase of the algorithm

searches for the transformation that brings the tem-

plate skeleton joints to their corresponding locations

in the scan data. These are detected by either us-

ing markers on the body (Allen,2003) or manually

(Seo, 2003). Local mapping is similar to that in (Ju,

2000; Ju, 2001) yet with more iterated relaxation and

re-mapping to maintain the surface regularity of the

generic model. The body of research work undertaken

in the area of conformation-based modelling has been

fruitful and has permitted achieving innovative appli-

cations such as the online garment design system de-

scribed in ( Cordier, 2003).

Certain other approaches have been developed re-

cently and are worth mentioning. Wang (Wang, 2005)

proposed a parametric model based on a particular

continuous surface model to which is added a layer

encapsulating speciﬁc details of shape. Ben Azouz et

al (Ben Azzouz, 2004; Ben Azzouz, 2005) proposed

a volumetric representation, wherein the 3D scan is

aligned inside a volume of ﬁxed dimension, and sam-

pled to a set of voxels (a unit of volume). An array of

signed distance between the voxels and their nearest

point in the scan is then derived and used as an HB

model. Principal component analysis is then applied

to such representation to extract the main types of

shape variation. This work has the merit of not rely-

ing on any anatomical landmarks. Victor and Paquet

Figure 6: Examples of models generated by interpolation

between two poses (Allen, 2002).

(Victor, 2005) used a representation that consists of

an array of cords, where a cord is deﬁned as a vector

that goes from the centre of mass of the human body

to the centre of mass of a given triangle modelling its

surface. This obviously necessitates ﬁrstly deriving a

triangular mesh model from the the HB scan data.

3.3.2 Dynamic Human Body Modelling

In fact, the topic of dynamic body modelling was

studied prior to the appearance of human body scan-

ners. The two main approaches in use today are

anatomical modelling and example-based modelling.

In the ﬁrst, the goal is to work out as accurate a model

as possible encompassing the skeleton, muscles and

other interior body structures, as well as the surface

of the skin, and permitting a systematic change in skin

shape when the underlying structure moves. A repre-

sentative works of this category are (Chadwick, 1989;

Magnenat-Thalmann, 1990; Turner, 1993; Scheep-

ers, 1997; Wilhelms, 1997; Aubel, 2001). These ap-

proaches proved to be effective in simulating body

dynamics and complex collisions, however at the ex-

pense of the computational cost, as each frame re-

quires its own simulation. Rather than looking for

a complex model to synthesize body movements and

deformation, the second kind of dynamic body mod-

eling, example-based modeling, adopts a data-driven

approach. The theory behind this approach involves

generating models in different key poses. These poses

are then correlated to various degrees of freedom,

with well-deﬁned joint angle values. New poses are

generated by smoothly interpolating among these val-

ues using interpolation techniques. This type of ap-

proach seems to have been inspired by the standard

key-frame techniques used in 2D animation. Authors

started by using man-made human body shapes in

a variety of poses, with the same underlying mesh

structure to simplify the correspondence between ver-

tices in each pose (Lewis, 2000; Sloan, 2001). When

it came to real data acquired by HB scanners, these

VISAPP 2006 - IMAGE ANALYSIS

194

approaches met new challenges: namely, the regis-

tration between the scan data in different poses and

the presence of holes and gaps inferred by self occlu-

sions. Allen et al (Allen, 2002) addressed these prob-

lems by using markers on the subject to establish a

correspondence between different poses and utilising

hole-ﬁlling techniques to reconstruct complete sur-

faces. In the same vein, Mohr et al (Mohr, 2003) ex-

tended the skeleton model by adding ﬁctitious joints

to help simulate non-linear body shape deformation

whereas local deformation is achieved via a linear re-

gression technique that ﬁts the vertices to the desired

locations.

4 CONCLUSION

Despite being a relatively recent research area, the

amount of work and research done on 3D human body

analysis and modelling demonstrates the increasing

interest in this topic and its wide range of potential

applications.

From the above overview, it emerges that important

steps have been taken towards bridging the gap be-

tween HBS technology and its various applications.

Some studies have already established the successful

exploitation of HBS in garment design, anthropomet-

ric applications, as well as entertainment.

A number of challenges remain to be dealt with,

however. For applications involving body landmark

detection, efforts undertaken to date have done a good

job of extracting visible body landmarks which are

apparent to the naked eye and which are at the same

time reﬂected in the HBS. On the contrary, other land-

marks are not visible, yet remain quite useful in an-

thropometric studies and garment design. These can

only be detected by touch and compression, however,

and thus are not embodied in HB scan data. One

potential approach will be developing an HB Atlas

that relates visible HB landmarks to invisible ones by

means of geometrical formulas. For instance, shoul-

der points can be deﬁned by the intersection of body

contours and the lines bisecting the angle between the

arm and the shoulder. Thus, invisible landmarks can

be derived from visible ones once the latter are iden-

tiﬁed.

Recent work done on HB segmentation has ef-

fectively addressed serious problems and challenges

such as posture variance and the deﬁciencies inherent

in data gathered through scans. Yet much work re-

mains to be done, especially involving postures where

limbs join, for example, crossed legs, or arms touch-

ing the torso. Dealing with such cases requires dis-

cerning the contours of discontinuities between the

joined parts of the body. Differential geometric tech-

niques, topological analysis and explicit model ﬁtting

could be elements of a potential approach.

By providing real data, HBS has been of great help

for researchers in improving the quality of human

body models from the points of view of geometry,

appearance and animation, and especially within the

framework of ”modelling from examples”. Much re-

search has already been done on applications in the

clothing industry and various entertainment sectors.

Despite these advances, the dream of a universal pa-

rameterised (and controllable) model of the human

body embodying the full spectrum of human body

shapes still appears to be quite far off. Indeed, the

complexity of the human body and the diversity of

factors that deﬁne its shape seem to truly plague the

development of such a model. Combining anatomical

modelling and example-based modelling into a single

framework might be a step in the right direction.

The three main areas of research into HBS data,

namely, human body modelling, body landmark de-

tection, and HBS segmentation, are in fact quite com-

plementary. Studies undertaken on HBS segmenta-

tion, for example, seem to have pointed the way to-

ward solutions to important issues in the other areas

in question. For instance, segmentation provides an

initial labelling of the body that can be used to re-

duce the search space of body landmarks. In the area

of dynamic human body modelling, segmentation has

contributed to solving problems arising with data reg-

istration. In effect, by identifying body parts in differ-

ent poses, it permits important correspondences be-

tween the related data sets, thus releasing the process

from relying on manually-generated markers or set

points for establishing such correspondences. It is

probable that further collaboration and interaction be-

tween researchers in these three research areas will be

greatly beneﬁcial for all concerned.

REFERENCES

Paquette,S. (1996). 3D Scanning in Apparel Design and

Human Engineering, In Computer Graphics and Ap-

plications Vol.16,No.9, pp.11-15.

CAESAR: Civilian American and European Surface An-

thropometry Resource http://store.sae.org/caesar/

Cyberware: http://www.cyberware.com.

TC2: http://www.tc2.com

Geisen,G.R., Mason,C.P., V.L Houston,V.L (1996). Au-

tomatic Detection, Identiﬁcation and Registration of

Anatomical Landmarks From 3-D Laser Digitizer

Body Segment Scans, In Proc. 17th Conference on the

Engineering in Medicine and Biology, Vol.1, pp.403-

404, Canada.

Pargas,R.P., Staples,N.J., Davis,J.S.(1996). Automatic

Measurement Extraction for Apparel from a Three-

SEGMENTATION AND MODELLING OF FULL HUMAN BODY SHAPE FROM 3D SCAN DATA: A SURVEY

195

Dimensional Body Scan, In Optics and Lasers in En-

gineering, Vol.28, PT.2, pp.157-172.

Jones,P.R.M., Li,P., Brook-Wavell,K., West,G.M.(1995).

Format of human body modelling from 3-D body

scanning, In Int. Journal of Clothing Science, Vol.7,

No.1, pp.7-16.

Li,P. Jones,P.R.M.(1997). Automatic Editing and Curve-

ﬁtting of 3-D Surface Scan Data of the Human Body,

In Proc. Int. Conference on Recent Advances in 3-D

Digital Imaging and Modelling, pp.296-301, Ottawa,

Canada.

Nurre,J.H.(1997). Locating Landmarks on Human Body

Scan Data, In Proc. Int. Conference on Recent

Advances in 3-D Digital Imaging and Modelling,

pp.289-295, Ottawa, Canada.

Nurre,J.H., Connor,J., Lewark,E., Collier,J.(2000). On seg-

menting the three-dimensional scan data of human

body, In IEEE Trans. On Medical Imaging, Vol.18,

No. 8, pp.787-797.

Dekker,L., Khan,S., West,E., Buxton,B., Tre-

leaven,P.(1998). Models for Understanding the

3D Human Body Form, In Proc. IEEE Workshop on

Model-Based 3D Image Analysis, pp.65-74, Bombay,

India.

Douros,I., Dekker,L., Buxton,B.F.(1999). Reconstruc-

tion of the surface of the human body from 3D

scanner data using B-spline, In Proc. SPIE on

Three-Dimensional Image Capture and Applications,

pp.234-245 San Jose, California.

Wang,C.L., Chang,T.K., Yuen,M.(2003). From laser-

scanned to feature human model: a system based on

fuzzy logic concepet In Computer-Aided Desingn,

Vol.35, pp. 241-253.

Brunsman,M.A., Daane,H.M., Robinette.K.M.,(1997). Op-

timal Postures and Positioning for Human Body Scan-

ning In Int. Conference on Recent Advances in 3-D

Digital Imaging and Modelling, pp.266-273, Ottawa,

Canada.

Lin,M.(1999). Tracking Articulated Objects in Real-Time

Image Sequence In Proc. Int. Conference on Com-

puter Vision, Corfu, Greece, pp.648-653.

Paquet,E., Robinette,K., Rioux,M.(2000). Management

of three-dimensional and anthropometric databases:

Alexandria and Cleopatra Journal of Electronic Imag-

ing, Vol. 9, No. 4, pp.412-431.

Paquet,E., Rioux,M.(2001). Inﬂuence of Pose on 3-D

Shape Classiﬁcation: Part II In Proc. Conference of

Digital Human Modeling for Design and Engineer-

ing, Arlington, Virginia, USA. Also in TR. NRC

44196, Intstitute of Information Technolofy, National

Research Council, Canada.

Werghi,N., Xiao.Y.(2002). Posture Recognition and Seg-

mentation from 3D Human Body Scans In Proc Int.

Conference of 3D Data Processing Visualization and

Transmission, Padova, Italy, pp. 456-570.

Werghi,N.(2005). A Discriminative 3D Wavelet-based De-

scriptors: Application to the Recognition of Human

Body Postures In Pattern recognition letters , Vol.26,

no.5, pp.663-677.

Xiao,Y., Werghi,N., Siebert, P.(2003). A Topological Ap-

proach for Segmenting Human Body Shape. In Proc.

Int. Conference on Image Analysis and Processing,

pp.82-87. Mantova, Italy.

Xiao,Y., Werghi,N., Siebert, P.(2003). A Topological ap-

proach based on Discrete Reeb-Graph for the Segmen-

tation of Human Body Scans In Proc. IEEE Int. Con-

ference on 3D Imaging and modeling, pp.378-385, Al-

berta, Canada.

Xiao,Y., Werghi,N., Siebert, P.(2004). Topological Seg-

mentation of Discrete Human Body Shapes in Vari-

ous Postures Based on Geodesic Distance Proc. In

Int. Conference on Pattern Recognition, pp.131-135,

Cambridge, UK.

Werghi,N., Xiao,Y., Siebert,J.P.(2005). A functional-based

segmentation of human body scans in arbitrary pos-

tures, To appear in IEEE Transactions on Systems,

Man, and Cybernetics.

Marr,D., Nishihara,H.K.(1976). Representation and recog-

nition of the spatial organization of three dimensional

shapes In AI memo No. 377

, Cambridge, Mass.: MIT,

Artiﬁcial Intelligence Laboratory.

Pentland.A.(1990). Automatic Extraction of Deformable

Parts Models In Int. Journal of Computer Vision,

Vol.4, pp.107-126.

Terzopoulos,D. D. Metaxas,D.(1991). Dynamic 3D mod-

els with Local and Global Deformation: Deformable

Superquadrics, In IEEE Trans. PAMI Vol.13, No.7,

pp.703-714.

Borges,D.L., Fisher,R.B.(1993). Segmentation of 3D Ar-

ticulated Objects by Dynamic Grouping of Disconti-

nuities, In Proc. British Machine Vision Conference

pp.279-287, Surrey,UK.

Dion Jr.D., Laurendeau,D., Bergevin,R.(1997). Generalized

Cylinders Extraction in a Range Image, In Int. Con-

ference on Recent Advances in 3-D Digital Imaging

and Modelling, pp.296-299, Ottawa, Canada.

Ferrie,F. Lagarde,J., Whaite,P.(1993). Darboux Frames,

Snakes and Super-Quadrics: Geometry from the Bot-

tom Up, IEEE Trans. PAMI, Vol.15, No.8, pp.771-

784.

Trucco,E.(1991) Inferring convex subparts from slice data,

Pattern Recognition Letters, Vol.1, No.5, pp. 707-715.

Kass,M., Terzopoulos,D.(1988). SNAKES: Active contour

models, Int. Journal of Computer Vision, Vol.1, pp.

321-332.

Ashbrook,A., Fisher,R.B, Werghi,N., Robertson,C.(1999).

Construction of Articulated Models from Range Data,

In Proc. British Machine Vision Conference, Vol.1,

pp.183-192, Nottingham,UK.

Shen,J., Thalmann,D.(1995) Interactive Shape Design Us-

ing Metaballs and Splines, In Proc. of Implicit Sur-

faces’95,Grenoble, France, pp.187-196.

Matsuda,R., Nishita,T.(1999). Moeling and deformation

Method of Human Body Model Based Method on

VISAPP 2006 - IMAGE ANALYSIS

196

Range Data, In Proc. Int. Conference on Shape Mod-

eling and Applications, pp.218-221

Muraki,S.(1991). Volumetric Shape Description of Range

Data Using Blobby Models, Computer Graphics, Vol.

25, No. 4, pp.223-227.

Hilton,A., Beresford,D., Gentils,T., Smith,R.,

Sun,W.(1999). Virtualp People: Capturing hu-

man models to populate virtual worlds, In Proc.

Computer Animation, IEEE Press, pp.174-185.

Starck,J. and Hilton,A.(2001). Human Shape Estimation

in a Multi-Camera Studio, In Proc. British Machine

Vision Conference Manchester, pp.573-583.

Kakadiaris,I.A., Metaxas,D.(1998) Three-dimensional hu-

man body model acquisition from multiple views, Int.

Journal of Computer Vision, 30(3), pp.191-218.

Ju X., Werghi N. and Siebert J.P.(2000). Automatic seg-

mentation of 3D human body scans In Proc. IASTED

Int. Conf. on Computer Graphics and Imaging, Las

Vegas, USA.

Ju,X. Siebert,J.P.(2001). Conforming generic animatbale

model to 3D scanned data, In Int. Conference of Nu-

merisation 3D-Scanning, pp.35-41, Paris, France.

Ashbrook,A.P., Fisher,R.B., Robertson,C.,

Werghi,N.(1998). Finding Surface Correspon-

dance for Object Recognition and Registration

Using Pairwise Geometric Histograms, In Proc.

European Conference on Computer Vision, Friburg,

Germany,pp.674-686.

Allen,B., Curless,B., Popovic,Z.(2003). The space of

all body shapes: reconstruction and parameteriza-

tion from range scans, In Proc. SIGGRAPH’2003,

Addison-Wesley, pp.587594.

Seo,H., Magnenat-Thalmann,N.(2003). An Automatic

Modelling of Human Bodies from Sizing Parameters,

In Proc. ACM SIGGRAPH symposium on Interactive

3D Graphics, Montery, Canada, pp.234-242

Wang,C.C.L.(2005). Parametrization and parametric de-

sign of mannequins, Computer-Aided Design, Vol. 31,

No.1, pp.83-98.

Ben Azzouz,Z., Rioux,M., Shu,C., Lepage,R.(2004) Analy-

sis of Human Shape Variation Using Volumetric Tech-

niques, Proc. Int. Conference on Computer Animation

and Social Agents, Geneva, Switzerland, pp.197-206.

Ben Azzouz,Z., Shu,C., Rioux,M., Lepage,R.(2005). Ex-

tracting Main Modes of Human Shape Variation from

3-D Anthropometric Data, In Proc. Int. Conference on

3-D Digital Imaging and Modeling, Ottawa, Ontario,

Canada, pp.335-342.

Victor,H.L., Paquet,E.(2005). Anthropometric Calibration

of Virtual Mannequins through Cluster Analysis and

Content-based Retrieval of 3-D Body Scans, In IEEE

Conf. on Instrumentation and Measurement Technol-

ogy, Ottawa, Ontario, Canada. Also in TR. NRC

47439, Intstitute of Information Technolofy, National

Research Council, Canada.

Cordier,F., Seo,H., Magn7enat-Thalmann,N.(2003). Made-

to-measure technologies for an online clothing store,

Computer Graphics and Applications, IEEE Volume

23, Issue 1,pp.38-48.

Chadwick,J.E., Humann,D.R. Parent,R.E.(1989). Layered

construction for deformable animated characters In

Proc. SIGRAPH 89, pp.243-252.

Magnenat-Thalmann,N., Thalmann,D.(1990). Human

Body Deformations Using Joint-dependent Local Op-

erators and Finite-Element Theory, In Making Them

Move (N. Badler, BA Barsky, D. Zeltzer, eds), Morgan

Kaufmann, San Mateo, California, 1990, pp.243-262.

Turner,R., Thalmann,D.(1993). The Elastic Surface Layer

Model for Animated Character Construction, In Proc.

Int. Conference on Computer Graphics, Lausanne,

Switzerland, pp.399-412.

Wilhelms,J., Van Gelder,A.(1997). Anatomically Based

Modeling, In Proc. SIGGRAPH ’97, pp.173-180.

Scheepers,F., Parent,R. E., Carlson, W.E. May,S.F.(1997).

Anatomy-based modeling of the human musculature,

In Proc. SIGGRAPH ’97, pp.163-172.

Aubel,A., Tahlmann,D.(2001). Efﬁcient Muscle Shape De-

fromation, Deformable Avatars

, Kluwer Publications,

pp.132-142.

Lewis,J.P., Cordner,M.(2000). Pose space deformation: A

uniﬁed approach to shape interpolation and skelton-

driven deformation, In Proc. Siggraph’2000, pp.165-

e172.

Sloan,P.P.J., Rose,C. Cohen,M.F.(2001). Shape by example,

In Proc. Symposium of Interactive Graphics, pp.135-

144.

Allen B., Curless B., Popovic Z.(2002). Articulated

body deformation from range scan data, In Proc.

SIGGRAPH’2002, Addison-Wesley, Vol. 21, No.3,

pp.612-619.

Mohr,A., Gleicher,M.(2003). Building Efﬁcient, Accurate

Character Skins from Examples, In Proc. SIGGRAPH

’2003, pp.165-172.

SEGMENTATION AND MODELLING OF FULL HUMAN BODY SHAPE FROM 3D SCAN DATA: A SURVEY

197