See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224401122

Statistical body height estimation from a single image

Conference Paper · October 2008

DOI: 10.1109/AFGR.2008.4813453 · Source: IEEE Xplore

CITATIONS READS
51 643

2 authors, including:

Chiraz Benabdelkader
Ecole Nationale des Sciences de l'Informatique
27 PUBLICATIONS 1,466 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

FaceBots View project

Web analysis and crawler View project

All content following this page was uploaded by Chiraz Benabdelkader on 15 May 2014.

The user has requested enhancement of the downloaded file.

 Statistical Body Height Estimation from a Single Image

Chiraz BenAbdelkader Yaser Yacoob

New York Institute of Technology Institute of Advanced Computer Studies
Abu Dhabi, United Arab Emirates University of Maryland, College Park, MD USA
[email protected] [email protected]

Abstract points lies on this plane, such as in [1]. In the uncalibrated

case, a general approach for this problem consists of first
We address the problem of estimating a person’s body computing the length ratio q of the line segment with re-
height from a single uncalibrated image. The novelty of our spect to another line segment in the scene of known length
work lies in that we handle two difficult cases not previously L, called the reference length, and then simply obtaining
addressed in the literature: (i) the image contains no refer- the solution as the product (q · L). This approach is purely
ence length in the background scene to indicate absolute geometric and comprises a whole family of metrology tech-
scale, (ii) the image contains the upper body part only. In a niques [18, 9, 8, 14, 12]. However, it poses two main dif-
nutshell, our method combines well-known ideas from pro- ficulties in practice: (i) a reference length may not always
jective geometry and single-view metrology with prior prob- be available, and (ii) under the general (strong) perspective
abilistic/statistical knowledge of human anthropometry, in a camera model, the length ratio q can only be computed un-
Bayesian-like framework. The method is demonstrated with der certain circumstances, such as when the two line seg-
synthetic (randomly generated) data as well as a dataset of ments are coplanar and the vanishing line of their plane is
96 frontal images. known, or when the two line segments are collinear and the
vanishing point of their direction is known.
The uncalibrated scenario is what we are concerned with
1. Introduction in this paper. Recently, we have developed a novel visual
In this paper, we re-visit the problem of body height metrology technique that extends the above described ap-
(or stature) measurement from a single uncalibrated image. proach for this scenario in one important way: it obviates
Body height estimation of people in images and video has the need for a reference length [2]. This is achieved by
many important applications, as body height can be used incorporating certain probabilistic/statistical properties of
to identify individuals, either uniquely or partially. Ap- human anthropometry into the estimation process. In the
plications where the camera calibration parameters are un- present work, we apply this same technique to body height
available include forensics and detection/tracking of people measurement from whole body and upper body images of
from a moving camera or across multiple camera views. In people. In other words, the input to our estimation methods
forensic image analysis, body height can be used to rule consists of anthropometric ratios (which are measured from
out the possibility that a particular person is the same per- the image) and anthropometric statistics. Our main contri-
son in the image (i.e. screening and elimination of suspects) bution lies in that we handle two difficult cases not previ-
[4, 6, 18, 7]. In human detection/tracking applications, such ously addressed in the literature as far as body height esti-
as video surveillance and customer tracking, body height mation from uncalibrated images: (i) the image contains no
can be used to distinguish among a small set of tracked peo- information on absolute scale (namely a reference length),
ple in the scene [1, 16, 3, 15, 10]. and (ii) the image contains the upper body part only.
From a projective geometry point of view, this amounts The main premise of our body height estimation method
to the basic 3D reconstruction problem of recovering the can be explained in an information-theoretic sense as fol-
length of a 3D line segment based on its projection in a lows. Basically, since human body dimensions and propor-
2D image. In the calibrated case, this problem can be tions (i) vary within a relatively limited range and (ii) are
solved with additional knowledge of the segment’s direc- correlated, the projection of a person in an image must en-
tion and a plane or line passing through one of its endpoints code some information about the absolute scale of the im-
[1, 16, 13]. A typical scenario consists of the line segment aged scene. Therefore, we expect to be able to estimate the
being orthogonal to the ”ground plane”, and one of the end- person’s body dimensions, such as body height, to within

978-1-4244-2154-1/08/$25.00 2008 IE
accuracy better than a random guess but less than when ac- has a unique solution, which happens to be the right singular
curate/complete absolute scale information is available. vector, denoted vn , of C that corresponds to its smallest
singular value [19]. Thus, a set of approximate (non-exact)
2. Methods solutions for (2) is:

We start by presenting the visual metrology technique x = kvn (4)

that we developed recently [2] (Section 2.1), then we dis-
cuss how this technique can be applied to body height esti- where k is an arbitrary scalar. This is an infinite and un-
mation from full-body images (Section 2.2) and upper-body bounded solution space, with infinite candidate solutions.
images of people (Section 2.3). Finally, we discuss some Furthermore, since Cx2 = k 2 Cvn 2 = k 2 σn2 , the data
details about the implementation of these methods in Sec- fit (namely the sum of square errors) of the solutions is an
tion 2.4. increasing function of k, and the trivial solution has the best
fit. Thus, data fit by itself does not seem to be a sensible
2.1. Statistical Visual Metrology way for selecting a single ”best” solution from among these
candidate solutions.
We describe a novel metrology technique that simultane-
As mentioned earlier, this situation arises because funda-
ously estimates multiple, say n, 3D line segments (such as
mentally this is an underdetermined inverse problem; there
anthropometrics) from a single image under minimal cali-
exist infinite x’s that can generate the same input y. Hence
bration information, with n > 1. Let x be a vector contain-
what we need to do is impose additional constraints on x in
ing the lengths of n 3D line segments, and let y be a vector
order to narrow down the solution space. This is generally
containing m length ratios xi /xj where i = j. The goal is
called regularization. To this end, we shall assume x is a
to estimate x based on the observed (known) value of y.
random vector (equivalently each xi is a random variable),
From a purely algebraic point of view, this is an ill-posed
and formulate constraints based on the following statistical
inverse problem, as there are infinitely many vectors x that
properties of x: (i) range of variation of both x and y, and
could generate the same value of y (at the very least, if a
(ii) prior probability distribution of x.
solution x0 exists, then any scalar multiple of x0 is also a
A few definitions first. Let Li,i and Ui,i be respectively
solution). In the sequel, we will show that it is possible
the lower and upper α percentiles1 of the probability dis-
to find an approximate solution by introducing appropri-
tribution of xi . Similarly, let Li,j and Ui,j respectively be
ate prior statistical knowledge about x in a Bayesian-like
the lower and upper bounds of the quotient random variable
framework.
ρij = xi /xj , for any i = j. Clearly, with a sufficiently
Let I be the set of pairs (i, j) such that the ratio xi /xj
small α, the α percentiles act as de facto lower and upper
is included in y. Let rij be the observed value of xi /xj ,
bounds of the respective random variable. We also assume
for any (i, j) ∈ I. Each rij value provides a linear equality
that the prior probability of x, denoted Π(x), has a Gaussian
constraint in the unknowns xi and xj :
distribution N (μ, Σ), and we define the log prior function
(also called mahalanobis distance) as:
xi − rij xj = 0 (1)
Hence with m distinct length ratios we obtain m corre- (x) = (x − μ)T Σ−1 (x − μ) ∝ log Π(x) (5)
sponding linear equality constraints, which we express in
matrix form as a linear homogeneous system: Now, if the values of Li,j and Ui,j are known for all
(i, j) ∈ I, then the following set of inequality constraints
Cx = 0 (2) on x are obtained:

Barring numerical errors, rij and rji provide equivalent

equality constraints, and hence C has at most n(n − 1)/2 Li,i ≤ xi ≤ Ui,i for all i ∈ {1, 2, · · · , n} (6)
rows. We know from basic linear algebra that an exact non- xi
trivial solution for the linear system in (2) exists if and only Li,j ≤ ≤ Ui,j for all (i, j) ∈ I (7)
xj
if C is rank-deficient, i.e. rank(C) < n. Otherwise, when
rank(C) = n, a non-exact nontrivial solution can be ob- Furthermore, we can include a smoothing term in the cost
tained by solving the following minimization problem via functional of the minimization problem in (3) that favors
least squares estimation (LSE): solutions with higher prior probability, i.e. smaller (x).
1 Let X be a random variable with mean μ and standard deviation σ.
min Cx2 subject to x2 = 1 (3) The lower α percentile of X is the value xL such that Pr(X ≤ xL ) = α,
x
and its upper α percentile is the value xR such that Pr(X ≤ xR ) = 1−α.
The condition x2 = 1 serves both to avoid the trivial Furthermore, if X has a normal distribution, then xL = μ − t · σ and
solution and to guarantee uniqueness. This LSE problem xR = μ + t · σ, where t = χ−1 (1 − 2α).
By doing so, in effect we incorporate knowledge about the acromial height, (4) head to chin distance (or head length),
correlatedness of the x variables (in addition to knowledge (5) stomion to top of head distance, (6) subnasale to top of
about their range of variation). Intuitively, this knowledge head distance, (7) forehead to chin distance, (8) sellion to
includes things like a person taller than 1.80m is unlikely to chin distance, (9) biocular distance, i.e. between outer cor-
have a head length of less than 12cm. We thus re-formulate ners of the eyes, (10) bitragion distance. These anthropo-
our inverse problem of (2) as follows: metrics are illustrated in Fig. 1.
For the first four anthropometrics the person is assumed
 to be in upright standing pose—not slouching or leaning on
Li,i ≤ xi ≤ Ui,i
min[Cx2 + λ(x)] subject to Li,j ≤ xxji ≤ Ui,j
one side. Anthropomerics 1–8 are all vertical and collinear
x
along the midline of symmetry, and so we only need the ver-
(8) tical vanishing point to compute their pairwise ratios. An-
where λ is a regularization parameter that controls the rel- thropometrics 4–10 lie on the facial surface and are nearly
ative effect of the data term and smoothing term. It is easy coplanar, provided that the person is sufficiently far from the
to see that this is in effect a linearly constrained quadratic camera. Furthermore, when the head pose is nearly frontal
function minimization problem (also called quadratic pro- to the camera, the weak perspective model is a good approx-
gramming (QP)), since the inequalities in (7) can be ex- imation over the facial region, and in this case computation
pressed as linear inequalities in xi and xj . We use standard of the pairwise ratios of facial anthropometrics requires no
iterative QP techniques to solve this quadratic program (cur- calibration information. Currently we compute pairwise ra-
rently the quadprog Matlab function). tios of anthropomerics 1–8 and pairwise ratios of anthro-
Interestingly, assuming input errors are of the form: pomerics 4–10, hence a total of 39 ratios.
Cx = η, where η is Gaussian white noise, then our so- However, if the whole body and/or the facial plane are
lution of (3) corresponds to the Maximum Likelihood es- at an angle with respect to the camera, then the weak per-
timate (MLE), and the solution of (8) corresponds to the spective assumption no longer applies, and so computing
Bayesian maximum a posteriori (MAP) estimate of the in- pairwise ratios of the facial anthropometrics (4–10) requires
verse problem in (2). a lot more information than what is available in an uncali-
brated image. Consequently we only use anthropometrics
2.2. Body Height from a Whole-Body Image 1–8 for such images.
Given a whole-body photograph of a person, with the Finally, a word on the choice of anthropometrics. In the-
person in upright standing pose (not slouching or leaning), ory, there are many other possible anthropometrics that can
we will show that the metrology technique of Section 2.1 be used, in addition to or in place of the ones we use in
can be used to estimate his/her body height. Namely, the n this paper. However, some factors have limited this choice,
line segments consist of a set of anthropometrics including in particular: (i) ability to compute the anthropometric ra-
body height, and each anthropometric is the straight line tios from an uncalibrated image, (ii) availability of anthro-
distance between two visible and well-defined body land- pometric statistics, and (iii) accuracy of landmark localiza-
marks (canonical points). Furthermore, because the image tion in the image; for example, the bigonial width3 is a bad
is uncalibrated, these anthropometrics are chosen such that choice because the gonial is difficult to locate in an image,
their ratios can be measured from the image, namely: they even manually and even in high-resolution images. The
need to be either collinear or coplanar, and the vanishing same can be said about arm length, since a person’s arm
points/lines, if needed, can be readily measured from the is often times bent and/or occluded by clothing.
image (see Section 1).
The premise of this body height estimation method can 2.3. Body Height from an Upper-Body Image
be formulated in an information-theoretic sense as follows. We extend the method of Section 2.2 to handle images
Since human body dimensions and proportions (i) vary containing only the upper body part of the person. In a nut-
within a relatively limited range and (ii) are correlated, the shell, we first estimate some set of upper body anthropo-
projection of a person in an image must encode some infor- metrics using the same technique, and then we use a linear
mation about the absolute scale of the imaged scene. There- prediction model to compute (or predict) body height from
fore, we expect to be able to estimate the person’s body di- the estimated anthropometric values. The model parame-
mensions, such as body height, to within accuracy better ters are learned from a large set of randomly generated data
than a random guess but less than when accurate/complete
shoulder, subnasale: the point just below the nose, tragion: the point on the
absolute scale information is available. cartilaginous flaps in front of each earhole, stomion: the point at the center
We currently use the following set of ten anthropomet- of the mouth, sellion: at the deepest point of the nasal root depression (also
rics2 : (1) body height, (2) trapezius (or neck) height, (3) called nose bridge) [11, 17].
3 Bigonial (or mandible) width: the straight-line distance between the
2 trapezius: the point at the side of the neck, acromion: the tip of the left and right gonial landmarks—the corners of the jaw.
 top of head
eyes
head ears
length
Body
Height chin
neck
acromions

Neck
Height

Acromial
Height

base of
feet

Figure 1. Anthropometrics we estimate in this paper.

(Section 3.1). The main challenge we face with this ap- Figure 2. Landmark localization in frontal-view image. The blue
proach lies in that, aside from facial anthropometrics, there dots are points we mark manually; green dotted line is the medial
are pretty much no anthropometrics that are at once: eas- axis of symmetry of the person, which we estimate by fitting a line
ily obtainable from images, strongly correlated with body to some of those dots (see text)
height, and whose statistics are available (see discussion at
the end of Section 2.2). Currently we use the seven facial
anthropometrics from Section 2.2 (4–10), plus the acromion
to top of head distance. Because statistics for the latter an-
thropometric are not available in any well-known anthro- tics, namely the lower and upper bounds of both each an-
pometric surveys, we have instead derived them from the thropometric and each anthropometric ratio, and the mean
statistics of body height and acromial height. and covariance of all anthropometrics. The statistics were
all obtained from the seminal anthropometric survey in [5].
2.4. Implementation Details The details are given in [2]. Table 1 shows these statistics
for some anthropometrics, with α =1e-12. It is important
Localization of the body landmarks associated with the to note, however, that these statistics are categorized by
anthropometrics of concern is achieved as follows (Fig. 2). gender, i.e. separate statistics for males and females. Con-
For frontal-view images, we start by locating these 15 body sequently, the gender of the person in the image is assumed
landmarks in the image: top of the head, forehead, sub- to be known, which is an important consideration for auto-
nasale, stomion, chin, left and right corners of the eyes, left mated implementations of the method.
and right tragions, left and right trapezius points (neck), left
and right acromions, and left and right medial longitudinal
foot arches. This is currently done semi-automatically by
having the user select points in the image via an interac- Table 1. Some anthropometric statistics (in cm), where SD: stan-
tive Matlab interface. We then estimate the person’s medial dard deviation, UB: upper bound, LB: lower bound. Based on the
1988 US Army survey [5] and with α =2.9e-7 (which under the
axis (midline of symmetry) as the line passing through the
assumption of normality means the upper and lower bounds corre-
top of the head, midpoint of the two foot landmarks, and
spond to 5 standard deviations from the mean).
(if known) the vertical vanishing point. We refine the lo-
cations of the first five landmarks (top of head, forehead,
Males Females
subnasale, stomion, chin) by projecting them onto the me- Mean SD LB UB Mean SD LB UB
dial axis. For the left/right landmark pairs of the trapezius, Body height 175.5 6.68 142.2 209.0 162.9 6.36 131.1 194.7
Head length 23.2 .88 18.8 27.6 21.76 .85 17.5 26.0
neck, acromions, and foot arches, we compute the intersec- Bitragion 14.5 .60 11.5 17.5 13.64 .52 11.0 16.2
tion of the medial axis with the line segment joining each Interocular 10.2 .54 7.50 12.9 9.62 .50 7.12 12.1
landmark pair. This way, the anthropometrics 1–8 are all
collinear with the medial axis.
Finally, a word about the required anthropometric statis-
 true uncorrupted estimated
randomly generate x compute pairwise y x
head to chin distance, (5) stomion to top of head distance,
Metrology
anthropometric
vector
anthropometric
ratios
Technique (6) subnasale to top of head distance, (7) forehead to chin
distance, (8) sellion to chin distance, (9) biocular distance,
Figure 3. Methodology for (random) generation of synthetic an- i.e. between outer corners of the eyes, (10) bitragion dis-
thropometric data (namely y vectors) to be used in testing our an- tance, (11) acromion to top of head distance.
thropometric estimation method. The body height estimation error is computed as follows:
estimated value - true value. For the method
parameters, we use λ = 0.01 and α =2.9e-7 (i.e. upper and
lower bounds correspond to 5 standard deviations from the
mean). In order to investigate the effect of choice of the an-
thropometric set on estimation error, we tested the method
with various different sets. The results are given below in
Figures 5–8 in two forms: (a) the cumulative distribution
of absolute estimation error, and (b) the distribution of es-
timation error (as a boxplot). Furthermore, for comparison
purposes, we have included in each Figure the performance
of the baseline algorithm, i.e. the one that estimates the un-
known body height as the population mean (175.5cm for
males and 162.9cm for females–See Table 1).
As far as estimation from whole-body images, our
method is significantly better than the baseline algorithm for
Figure 4. Sample images from our inhouse dataset. synthetic data, but only slightly better for real data. Since
the synthetic vectors are uncorrupted by noise, the syn-
thetic estimation error actually corresponds to the method’s
3. Experiments and Results model error. This suggests that the method is quite sensi-
tive to input error in the real data, which mostly comes from
3.1. The Data landmark localization and vanishing point/line estimation.
We tested our anthropometric estimation method using Furthermore, generally speaking, the method’s performance
both synthetic data and real images. The former consists of improves when more anthropometrics are used in the esti-
randomly generated y vectors, each of which is obtained by mation. This especially obvious for synthetic data. As to
first randomly generating a vector x from the multivariate estimation from upper-body images, our method is no bet-
Gaussian distribution of male or female anthropometrics, ter than the baseline algorithm, both for synthetic and real
then computing the corresponding pairwise ratios (Fig. 3). data! This suggests that the main source of error lies in the
The real images consist of a set of high-resolution linear prediction model that we currently use.
(4368x2912) images captured in-house using a Canon 28-
200mm EOS camera (Fig. 4). Each image is a full-body 4. Conclusions and Future Work
shot of one person. The dataset contains a total 108 images,
We presented methods for estimating a person’s body
with 96 frontal-view shots. Also, the dataset comprises 27
height from a single uncalibrated image. The novelty is
different adults, 7 females and 20 males, spanning various
two-fold: (i) we do not require scale information (the so-
ethnicities: caucasian, chinese, indian, and african. The ver-
called reference length); and (ii) we handle images contain-
tical vanishing point needed to compute pairwise ratios of
ing the upper-body part only of the person. The method
anthropometrics 1–8 is computed as the intersection in the
was tested both synthetic data and real images. As far as
image of parallel vertical lines of the background scene.
whole-body images, performance is slightly better than the
baseline algorithm (that randomly guesses the population
3.2. Results
mean), but inferior to that of other well-known single-view
As discussed, we estimate body height by applying our metrology techniques such as [9]. This is expected since
novel metrology technique (Section 2.1) to estimate a set of our method uses less input information. Furthermore, since
anthropometrics. For whole-body images, this set includes quite a bit of estimation error is caused by input error in
body height (Section 2.2), while for upper-body images it the localization of landmarks and computation of vanishing
does not, and a linear model is subsequently applied to pre- points, this error can be reduced by estimating from multi-
dict body height (Section 2.3). To simplify the discussion, ple images or a video sequence. As for upper-body images,
let us assign a number to each anthropometric, as follows: our method is currently no better than the baseline algo-
(1) body height, (2) neck height, (3) acromial height, (4) rithm. However, there is room for improvement by using
 1 1
Cumulative distribution of body height estimation error

Cumulative distribution of body height estimation error

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3
baseline baseline
0.2 1−3 0.2 1−3
1−3,11 1−3,11
0.1 1−8 0.1 1−8
1−10 1−10
0 0
0 5 10 15 20 25 0 5 10 15 20 25 30
absolute body height error (cm) absolute body height error (cm)

(a) (a)

20 15

15 10

10 5
body height error (cm)

body height error (cm)

5 0

0 −5

−5 −10

−10 −15

−15 −20

−20 −25
−25 −30
baseline {1−3} {1−3,11} {1−8} {1−10}
baseline {1−3} {1−3,11} {1−8} {1−10}

(b) (b)
Figure 5. Synthetic data results for body height estimation from a Figure 6. Real data results for body height estimation from a
whole-body image. whole-body image.

a better prediction model than the naive linear model we [3] A. Bovyrin and K. Rodyushkin. Human height prediction
currently use. In future work, we plan to: (i) explore other and roads estimation. In IEEE Conference on Advanced
prediction models, particularly non-linear ones, for estimat- Video and Signal-Based Surveillance, 2005.
ing body height from upper-body images; (ii) extend the [4] S. Bramble, D. Compton, and L. Klasen. Forensic im-
method to estimation from multiple images; and (iii) inves- age analysis. In 13th Interpol Forensic Science Symposium,
tigate the effect of camera viewpoint (i.e. camera angle with 2001.
respect to person) on estimation accuracy. [5] J. Cheverud, C. G. Gordon, R. Walker, C. Jacquish, L. Kohn,
A. Moore, and N. Yamashita. 1988 anthropometric survey of
us army personnel. Technical Report TR-90/032, US Army
References Natick Research Labs, May 1990.
[1] C. BenAbdelkader, R. G. Cutler, and L. S. Davis. View- [6] D. Compton, C. Prance, M. Shears, and C. Champod. A
invariant estimation of height and stride for gait recognition. systematic approach to height interpretation from images. In
In Workshop on Biometric Authentication, 2002. Proc. of SPIE, volume 4232, 2001.
[2] C. BenAbdelkader and Y. Yacoob. Computational Forensics, [7] A. Criminisi, I. Reid, and A. Zisserman. Single view metrol-
volume 1, chapter Statistical Estimation of Human Anthro- ogy. In CVPR, pages 434–442, Sept. 1999.
pometry from a Single Uncalibrated Image, pages 200–220. [8] A. Criminisi, I. Reid, and A. Zisserman. Single view metrol-
Springer Verlag, 2008. ogy. IJCV, 40(2):123–148, Nov. 2000.
 1 1
Cumulative distribution of body height estimation error

Cumulative distribution of body height estimation error

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6

0.5 0.5

0.4 0.4

0.3 0.3
baseline baseline
0.2 4−8 0.2 4−8
9−10 9−10
0.1 4−10 0.1 4−10
4−11 4−11
0 0
0 5 10 15 20 25 0 5 10 15 20 25 30 35
absolute body height error (cm) absolute body height error (cm)

(a) (a)

20 15

10
15
5
10
body height error (cm)

body height error (cm)

0
5
−5
0
−10
−5
−15
−10
−20
−15
−25
−20
−30

baseline {4−8} {9−10} {4−10} {4−11} baseline {4−8} {9−10} {4−10} {4−11}

(b) (b)
Figure 7. Synthetic data results for body height estimation from an Figure 8. Real data results for body height estimation from an
upper-body image. upper-body image.

[9] A. Criminisi, A. Zisserman, L. Van Gool, S. Bramble, and Computer Vision. Cambridge University Press, second edi-
D. Compton. A new approach to obtain height measurements tion, 2004.
from video. In Proc. of SPIE, volume 3576, 1998. [15] A. Leykin and M. . Tuceryan. A vision system for auto-
[10] D. DeAngelis, R. Sala, A. Cantatore, P. Poppa, M. Dufour, mated customer tracking for marketing analysis: low-level
M. Grandi, and C. Cattaneo. New method for height es- feature extraction. Technical Report TR612, Indiana Univer-
timation of subjects represented in photograms taken from sity, June 2005.
video surveillance systems. International Journal of Legal [16] C. Madden and M. Piccardi. Height measurement as a
Medicine, 121(6):489–492, November 2007. session-based biometric. In Image and Vision Computing
[11] L. G. Farkas, editor. Anthropometry of the Head and Face. New Zealand, 2005.
Raven Press, second edition, 1994. [17] S. Pheasant. Bodyspace: Anthropometry, Ergonomics and
[12] F. Guo. Calibration and Metrology Using Still and Video the Design of Work. Taylor and Francis, second edition,
Images. PhD thesis, University of Maryland, College Park, 1996.
2007. [18] N. Saitoh, K. Kurosawa, and K. Kuroki. A study on height
[13] D. Hansen, B. Mortensen, P. Duizer, J. Andersen, and measurement from a single view. In ICIP, 1999.
T. Moeslund. Automatic annotation of humans in surveil- [19] G. Strang. Introduction to Linear Algebra. Wellesley-
lance video. In Fourth Canadian Conference on Computer Cambridge Press, third edition, 2003.
and Robot Vision, pages 473–480, 2007.
[14] R. Hartley and A. Zisserman. Multiple View Geometry in

View publication stats