Li S.Z., Jain A.K. (eds.) Encyclopedia of Biometrics

Подождите немного. Документ загружается.

¼ððI

þ I

ÞI

Þð1Þ

Afterwards, a Sobel edge-detection is performed,

obtaining a binary image where only the borders are in

black, while the rest of the image is in white. With this

result, the way to obtain the 31 absolute features is by

locating singular points in the image, and counting

pixels among them. In [5], authors proposed the fol-

lowing basic features (as seen in 3d):



Palm width at a certain point (avoiding conﬂict

with pegs). It is obtained as the number of pixels

from the ﬁrst black pixel on the right side of the top

view of the hand, to the next black pixel along the

same hor izontal line, after going through a set of

white pixels moving to the left.



Finger w idths at certain points, also avoiding pegs

and ring area, and obtained in an analogue way as

mentioned with the palm width.



Palm and ﬁnger heights, through the same mecha-

nism as mentioned above, but this time with the

side view of the hand.



Finger curvature or Deviations. This is deﬁned by

the distance between a middle point of the ﬁnger

and the middle point of the straight line between

the inter-ﬁnger point and the last height where the

ﬁnger width is measured. Equation used can be

seen below, where exponents refer to the coordi-

nate used, and subindex deﬁnes the ﬁnger used and

the width measurement used.



Angles between inter-ﬁnger points and the hori-

zontal line, which reﬂects the depth of each of the

inter-ﬁnger point.

deviation ¼ P



 P

ðP

 P

Þð2Þ

From those features, the feature space was grown

by adding relative measurements, i.e., relationships

among different sets of basic features. Authors, after

applying a Principal Component Analysis, discovered

that from all those measurements, only 25 features had

signiﬁcant discriminant properties. Figure 3d shows

the 25 absolute measurements from whose the ﬁnal

25 features extracted.

Authors researched the behavior of four different

comparators: Euclidean Distance, Hamming Distance

in the continu ous domai n (as seen in eq uation (3)

where x

refers to the ith component of the sample, L

is the feature vector length, t

is the mean of the ith

component, and t

is the standard of the ith compo-

nent), Gaussian Mixture Models (GMM) and Radial

Basis Function Neural Networks (RBF-NN). Also they

analyzed the dependence of the performance with the

number of samples used during enrollment. Results

showed that best performance was achieved with

GMMs, using ﬁve enrollment samples, and that the

system did not loose much of the performance is the

number of features reduced dow n to 15 (another rele-

vant work can be seen in [6]). As the number of

features is so low and each of them can even be

coded in one single byte, the viability of integrating

this modality with smart cards was a reality, and even

the development of a match-on-card prototype was

shown in [7].

hamming

ðx

; t

Þ¼# i 2f1; :::; Lg=jx

 t

j > t



ð3Þ

As already mentioned, it is of signiﬁcant impor-

tance to show the success of the ﬁrst commercial sys-

tems, because these systems demonstrated the viability

of this biometric modality in real scenarios. The ﬁrst

unit shown was in 1972 from Identimat, but popularity

was gained by the products of Recognition Systems.

They developed their ﬁrst prototype named HandKey

ID3D before 1990, and improved such system in 1997

by launching HandKey II (shown in Fig. 1). Hundreds

of thousand units have been sold, including applica-

tions in Universities, Airports, or Nuclear Plants. This

technology has gained wide application and accep-

tance, especially in Access Control Systems, and in

Time and Attendance Control.

Evolutions from Initial Works

From the results shown in the previous mentioned

works, several R&D groups have worked in this biomet-

ric modality. They have improved the system in several

ways. One of those working lines has been improving

usability by removing the orientation pegs. Some

researchers use a commercial scanner (e.g., [8, 9]),

while others have worked not only in a peg-free, but

also a contact-free system (e.g., [6, 10, 11]).

Other working lines have been focused in new

feature extraction approaches. Some authors have

increased the number of features, by including not

only geometrical measurements, but also information

680

Hand Geometry

about the hand contour [12]. Kumar and Zhang [13]

improved veriﬁcation rates in 4–7%, by discretizazing

features based on entropy studies. Gross et al. [9] have

worked with Active Appearance Models. Others have

worked in modeling hand contour and extracting fea-

tures by curvature gradient (e.g., [14]). Ma et al. model

hand geometry by using B-Spline curves [15]. Other

authors work with neural networks (e.g., [12]), either

for performing the whole identiﬁcation process, or just

for the comparison block.

Most of these studies claim error rates below 5%.

Some authors give even better ﬁgures, approaching a

99% of identiﬁcation accuracy. But even though, there

are some major open issues regarding this biometric

modality. One of those is the size of databases used for

testing. Unfortunately in most works such databases

are quite small, going up to 100 users with 10 photos

per user.

There are still some open issues, especially nowa-

days and within some kind of population. The expan-

sion in the use of jewelery, such as rings or piercings

with a wide variety of shapes and sizes, can be consid-

ered as image artifacts by hand recognition systems,

and lower the identiﬁcation rates. Also tattoos, and

specially those made in several colors can provoke the

denial of use by the hand recognition system. These

kind of problems have to be considered by new sys-

tems, to gain universality.

Usability and Multimodality

One of the most important facts related to this bio-

metric modality is its great usability. It seems that users

do not feel themselves afraid of using the system,

neither of noting their privacy attacked. Kukula and

Elliott [16] carried on a study that showed that 93% of

users enjoyed the system, nearly all found it easy to use,

and no one had privacy concerns. Kukula et al. [17]

have also studied the effects of training and habitua-

tion in using the system, show ing a better performance

when users are familiar with the identiﬁcation dev ice.

This great usability, together w ith the fact that

other biometric modalities use the same part of the

body (e.g., palmprints or ﬁngerprints), have pushed

researchers based on mult imodal biometrics to use this

biometric modality. Fusion works using palmprints

and hand geometry can be found in [18]or[8].

Other authors work even with three modalities,

adding ﬁngerprints to the previously mentioned ones,

like in [19]or[9]. Or even some authors have devel-

oped multimodal prototypes with other non-hand-

based modalities [20, 21].

Summary

Hand Geometry is a biometric modality whose pro-

mising features are the ease of use and high friendlines s

to the user. Furthermore, researchers have demon-

strated that error rates below 5% are possible, and

when applied to limited number of users, the level of

performance is high enough for certain applications.

Commercial products have found their business appli-

cations in Access Control Systems, as well as in Time

and Attendance environments.

Related Entries

▶ Gaussian Mixture Models

▶ Hand Data Interchange Format

▶ Hand Databases and Evaluation

▶ Hand-Geometry Device

▶ Match-on-Card

▶ Multibiometrics

▶ User Acceptance

References

1. Schlage, R.S.: Main website. http://recognitionsystems.schlage.

com/

2. Sidlauskas, D., Tamer, S.: In: Handbook of Biometrics, chap.

Hand geometry recognition, pp. 91–107. Springer, Berlin (2008)

3. Golfarelli, M., Maio, D., Maltoni, D.: On the error-reject trade-

off in biometric veriﬁcation systems. IEEE Trans. Pattern Analy.

Mach. Intell. 19(7), 786–796 (1997)

4. Jain, A., Ross, A., Pankanti, S.: A prototype hand geometry-

based veriﬁcation system. In: Second International Conference

on Audio and Video-Based Biometric Person Authentication

(AVBPA), pp. 166–171 (1999)

5. Sanchez-Reillo, R., Sanchez-Avila, C., Gonzalez-Marcos, A.: Bio-

metric identiﬁcation through hand geometry measurements.

Trans. Pattern Analy. Mach. Intell. 22(10), 1168–1171 (2000).

DOI 10.1109/34.879796

6. Kumar, A., Zhang, D.: Personal recognition using hand shape and

texture. IEEE Trans on Image Process. 15, 2454–2461 (2006)

7. Sanchez-Reillo, R., Gonzalez-Marcos, A.: Access control system

with hand geometry veriﬁcation and smart cards. IEEE Aero-

space Electr. Syst. Mag. 15(2), 45–48 (2000)

Hand Geometry

681

8. Savic, T., Pavesic, N.: Personal recognition based on an image of

the palmar surface of the hand. Pattern Recogn. 40(11),

3152–3163 (2007)

9. Ferrer, M., Morales, A., Travieso, C., Alonso, J.: Low cost

multimodal biometric identiﬁcation system based on hand

geometry, palm and ﬁnger print texture. In: 2007 41st

Annual IEEE International Carnahan Conference on

Security Technology, pp. 52–58 (2007). DOI 10.1109/

CCST.2007.4373467

10. Gross, R., Li, Y., Sweeney, L., Jiang, X., Xu, W., Yurovsky, D.:

Robust hand geometry measurements for person identiﬁca-

tion using active appearance models. In: First IEEE Interna-

tional Conference on Biometrics: Theory, Applications, and

Systems, 2007. BTAS 2007, pp. 1–6 (2007). DOI 10.1109/

BTAS.2007.4401936

11. Zheng, G., Wang, C.J., Boult, T.E.: Application of projective

invariants in hand geometry biometrics. IEEE Trans. Inform.

Forens. Secur. 2(4), 758–768 (2007) DOI 10.1109/TIFS.

2007.908239

12. Faundez-Zanuy, M.: Biometric veriﬁcation of humans by means

of hand geometr y. In: 39th Annual 2005 International Carnahan

Conference on Security Technology, CCST ’05, pp. 61–67

(2005). DOI 10.1109/CCST.2005.1594816

13. Kumar, A., Zhang, D.: Hand-geometry recognition using entro-

py-based discretization. IEEE Trans. Inform. Forens. Secur. 2(2),

181–187 (2007). DOI 10.1109/TIFS.2007.896915

14. Boreki, G., Zimmer, A.: Hand geometry: a new approach for

feature extraction. In: Fourth IEEE Workshop on Automatic

Identiﬁcation Advanced Technologies, pp. 149–154 (2005).

DOI 10.1109/AUTOID.2005.33

15. Ma, Y., Pollick, F., Hewitt, W.: Using b-spline curves for hand

recognition. In: Proceedings of the 17th International Confer-

ence on Pattern Recognition, 2004. ICPR 2004, 3, 274–277,

Vol. 3 (2004). DOI 10.1109/ICPR.2004.1334520

16. Kukula, E., Elliott, S.: Implementation of hand geometry

at purdue university’s recreational center: an analysis of

user perspectives and system performance. In: 39th Annual

2005 International Carnahan Conference on Security Technolo-

gy, 2005, CCST ’05, pp. 83–88 (2005). DOI 10.1109/

CCST.2005.1594879

17. Kukula, E.P., Gresock, B.P., Elliott, S.J., Dunning, N.W.:

Deﬁning habituation using hand geometry. In: 2007 IEEE

Workshop on Automatic Identiﬁcation Advanced Techno-

logies, pp. 242–246 (2007). DOI 10.1109/AUTOID.2007.

380627

18. Kumar, A., Wong, D.C., Shen, H.C., Jain, A.K.: Personal authen-

tication using hand images. Pattern Recogn. Lett. 27(13),

1478–1486 (2006)

19. Yang, F., Ma, B., Wang, Q.X., Yao, D., Fang, C., Zhao, S.,

Zhou, X.: Information fusion of biometrics based-on ﬁnger-

print, hand-geometry and palm-print. In: 2007 IEEE Workshop

on Automatic Identiﬁcation Advanced Technologies, pp. 247–

252 (2007). DOI 10.1109/AUTOID.2007.380628

20. Ross, A., Jain, A.: Information fusion in biometrics. Pattern

Recogn. Lett. 24(13), 2115–2125 (2003)

21. Jain, A., Nandakumar, K., Ross, A.: Score normalization in mul-

timodal biometric systems. Pattern Recogn. 38(12), 2270–2285

(2005)

Hand Geometry View Record – HGVR

Block of data that contains a hand silhouette captured

from one camera point of v iew during one hand

placement.

▶ Hand Data Interchange Format, Standardization

Hand Physiology

▶ Anatomy of Hand

Hand Shape

NICOLAE DUTA

Nuance Communications, Burlington, MA, USA

Synonym

Hand contour

Definition

A hand shape biometric system uses a camera or

scanner-based device to acquire the hand image of a

person from which shape information is extracted and

compared against the information stored in a database

to establish identity. Due to its limited discrimination

power, a hand shape biometric system mostly operates

in veriﬁcation mode; that is the system conﬁrms or

negates the claimed identity of an individual.

Introduction

An increasing number of systems require positive iden-

tiﬁcation before allowing an individual to use their

services. Biometric systems are already employed in

domains that require some sort of user veriﬁcation.

It is generally accepted that ﬁngerprint and iris pat-

terns can uniquely deﬁne each member of an extremely

682

Hand Geometry View Record – HGVR

large population which makes them suitable for large-

scale recognition (establishing a subject’s identity).

However, in many small-population applications, be-

cause of privacy or limited resources, it is only needed

to authenticate a person (conﬁrm or deny the person’s

claimed identity). In these situations, traits with less

discriminating power such as hand shape, hand geom-

etry, voice or signature can be used.

As noted in [1], hand shape-based authentication is

attractive due to the following reasons:

1. Hand shape can be captured in a relatively user

convenient, nonintrusive manner by using inex-

pensive cameras.

2. Extracting the hand shape information requires

only low resolution images, and the user templates

can be efﬁciently stored (120-byte templates are

reported in [1]).

3. This biometric modality is more acceptable to the

public mainly because, it lacks criminal connotation.

4. Additional biometric features such as hand geome-

try, palmprints, and ﬁngerprints can be easily

integrated to an existing hand shape-based authen-

tication system.

Operation of a Hand Shape-Based

Biometric System

A hand shape-based biometric system operates accord-

ing to the general diagram in [2] Fig. 2. In the enroll-

ment stage, hand shape data is acquired from the

registered users, feature sets are extracted from the

acquired data, and one or multiple templates per indi-

vidual are computed and stored in a database. In the

deployment stage, one snapshot of the user’s hand is

captured; a feature set is computed and then compared

to the user’s templates in the database. Based on the

comparison result, the claimed identity is accepted

or denied. As described in [2], the system comprises

the following modules: the sensor module, the feature

extraction module, the matching module, and the

decision-making module.

The sensor is usually a low/medium resolution

CCD camera attached (beneath or above) to a platform

on which the hand is placed (Fig. 1a ). Some multi-

modal biometric systems capture the palm surface

which includes both the hand shape and palmprints

[3]. Other systems capture the dorsal surface of the

hand from which only the hand contour can be

extracted (Fig. 2,[1, 4, 5]). Some of the systems in-

clude on the platform 4–6 pegs to guide the placement

of the user’s hand [4, 5]. Several researchers noted that

the guidance pegs deform the hand contour and de-

crease user convenience and proposed peg-less setups

[1, 3, 6]. In a few systems, the sensor consisted of a 45

dots per inch (DPI) scanner [3, 6].

In the feature extraction module, a set of discrimi-

nating features is computed from a user’s raw hand

image(s). The hand images are ﬁrst pre-processed in

order to extract the hand contour and eliminate arti-

facts such as the guidance pegs, user rings, overlapping

cuffs, or creases around the contour due to too light or

Hand Shape. Figure 1 (a) Example of a hand shape image acquisition system. (b) ROC curves for a hand shape-based

veriﬁcation system. The three curves correspond to feature vectors extracted from three, four and ﬁve ﬁngers.

Hand Shape

683

too heavy hand pressing. The pre-processing step can

range from simple

▶ thresholding [1, 5] to sophisticate

gray-level segmentation (

▶ Image segmentation)[4].

Possible dents at the artifact location are smoothed by

linear interpolation [4, 5] and/or morphologic opera-

tors [3].

In order to properly compare feature vectors

extracted from hand images, on e has to align the hand

contours such that each feature is computed from the

same region of the hand. Most of the older systems

relied on the pegs to align the hand images. However,

if the user is untrained or does not cooperate to proper

use of the hand scanner, then the resulting images are

not aligned (Fig. 2(c)) and the system’s veriﬁcation

performance degrades [4]. Therefore, it is necessary

to automatically align the acquired hand shapes before

extracting the feature vectors used for veriﬁcation. Due

to the ﬂexible nature of t he palm and ﬁngers, there may

be no linear transformation which accurately aligns

two hand contours. Hence, many of the proposed

alignment procedures detect and align each ﬁnger sep-

arately. The simplest ﬁnger alignment method consists

in registering the ﬁngertip, the two adjacent valley

points and several equally spaced points along the

contour in between the three landmarks [5]. Similarly,

a translation, rotation, and scaling can be found to

align the ﬁnger with symmetr y axis [6]. A more so-

phisticated alignment procedure ( based on quasi-

exhaustive polynomial search of point pair matching

between two sets of contour points) is presented in [4].

This procedure has the advantage of always ﬁnding a

good alignment even if the valley-point landmarks are

not accurately detected. The alignment step can be

avoided if the set of features extracted from the hand

image is invariant to Euclidean transformations [1].

The hand shape can be modeled either explicitly as

a set of 2D coordinates of several landmark points

along the hand contour [4, 5] or implicitly as a binary

image of the hand over an empty background [1, 6].

The two representations are intrinsically equivalent;

each of them can be easily derived from the other.

With both representations, dimensionality reduction

procedures may have to be applied as the original data

typically has a high dimensionality (see the ﬁfth col-

umn in Table 1). The dimensionality reduction meth-

ods most used are

▶ principal component analysis

(PCA) and independent component analysis (ICA),

and are applied to either the original data or to a

transformed version of the data (e.g., the Zernike

moments in [1]).

One or several templates per user may be created

during the enrollment stage and stored in the system’s

database. The templates are either the raw feature

vectors computed from a user’s hand images or the

average of those feature vectors.

The matching module compares a user feature

vector against the user’s template(s) stored in the da-

tabase in order to generate matching scores. Since the

feature vectors are usually points in an N-dimensional

Euclidean space, any metric distance can be used for

computing a matching score: Euclidean distance [1],

▶ Mahalanobis distance, absolute (L

) distance [6],

correlation coefﬁcient, etc. A few studies explicitly

model the class-conditional probabilities under Gauss-

ian assumptions [5]. As an exception, [3, 4]

▶ Procrus-

tes shape distance can be used since the feature vectors

are shapes corresponding to the hand contour. The

matching score is a positive number which shows the

dissimilarit y between the user’s hand and the templates

in the database.

Hand Shape. Figure 2 Hand shape alignment. Two scans of the same hand: (a–b) original images, (c) hand shapes

extracted from (a) and (b) overlaid, and (d) ﬁnger aligned shapes (Mean alignment error = 2.20 pixels).

684

Hand Shape

The ﬁnal decision concerning the user’s identity

(identiﬁcation) or the user’s claimed identity (veriﬁca-

tion) is taken by the decision module. In veriﬁcation

mode, the decision is typically threshold based: if the

matching score is below a given threshold the claimed

identity is accepted, otherwise it is rejected. The

threshold value is chosen based on the system’s ROC

curve such that the system satisﬁes some operating

constraints (e.g., an upper bound on the false accept

rate, an equal error rate, etc.). In identiﬁcation mode,

the incoming feature vector is typically assigned the

identity of the closest database template if the distance

to that closest template is lower than the veriﬁcation

threshold, otherwise the feature vector is considered

to belong to an imposter.

Performance Evaluation

For most of the research systems, performance evalua-

tion can only be based on the results and comparisons

provided by their authors. Several enrollment and per-

formance evaluation methodologies are discussed in

[5]. If extensive enrollment data can be acquired, best

performance is attained when an average template is

computed from each user’s enrollment measurements.

However, such a system is less user-friendly and more

difﬁcult to deploy. A more realistic dep loyment scenar-

io requires only one or few enrollment measurements

per user. In such case, a template may actually be a raw

feature vector and the main system parameter to

estimate is the decision threshold. Most researchers

split the available measurements for each user into an

enrollment set and a testing set, and evaluate the

threshold value based on the enrollment data. This

has an advantage that the training data is representa-

tive for the test data, i.e., one expects to obtain an

estimate of the decision threshold which works as

well on the test data. In commercial deployments

though, the system may be trained by the manufactur-

er while the enrollment is performed by the customer

who has to use the factory set threshold.

It is difﬁcult to directly compare the performance

ﬁgures reported in the literature. The main reason is the

absence of (1) a common benchmark data set and (2)

standard enrollment and testing procedures. The differ-

ent datasets used for research introduce several variation

factors in the systems reported: (1) population size,

(2) population age and/or structure, and (3) users’ mo-

tivation to cooperate. Table 1 summarizes the perfor-

mance of some research systems tested in identity

veriﬁcation mode. Some authors only report equal

error rate (EER) ﬁgures while others include the sys-

tem ROC curve. When the ROC cur ve was present, the

estimated FRR rate corresponding to FAR = 0.01. As it

can be seen in the last column of Table 1, most systems

reported error rates on the order of 10

2

. A few sys-

tems have also been tested in identiﬁcation mod e

and report identity recognition errors of 1–3% [6].

Note that no separate imposter population is used so

the recognition error may be under estimated. Some

authors integrate hand shape features with palm-based

Hand Shape. Table 1 Comparison of some hand shape-based systems presented in the literature

System

Population

size

Samples/

person

Number/

type of

templates Features used

Similarity

measure Performance

[1] 40 10 5 (raw

feature

vectors)

Zernike moments of the binary

hand image followed by PCA

(30)

Euclidian FAR = 0.01

FRR = 0.02

EER = 0.0164

[2] 53 2–15 1–14 (raw

contours)

Hand contour coordinates

(120–350 contour points)

Mean alignment

error

FAR = 0.01

FRR = 0.06

[4] 51 10–20 1 (average) +

multiple raw

Hand contour coordinates,

angles (51–211 contour points)

Log-Likelihood

under Gaussian

assumption

EER = 0.00001 –

0.002

[5] 458 3 2 (raw

feature

vectors)

Contour coordinates (2048

points) ICA on binary hand

image (458)

Modified

Hausdorff L1,

cosine distance

EER = 0.01– 0. 02

Hand Shape

685

features (which can be acquired through a single image

measurement) and report veriﬁcation error rates on

the order of 10

3

[3].

Limitations of the Hand Shape-Based

Biometric Systems

There are two factors which make a hand shape-based

biometric system less accurate than a ﬁngerprint or

iris-based system [4]:

1. The hand shape is not unique within a large popu-

lation. That is demonstrated in Fig. 3: after ﬁnger-

alignment, the hand contours of two different

users are almost identical. Therefore any geometric

features extr acted from the two aligned contours

will be very similar and the system will likely con-

fuse the identity of the two persons. In [4] three

pairs of users with very similar hand shapes

within a population of 53 persons were iden-

tiﬁed. However, the problem is alleviated if the

system is used in veriﬁcation mode since an impo-

ster is less likely to know the identity of the

registered user(s) whose hand shape best matches

his or hers.

2. The human hand is a ﬂexible object and its contour

may suffer non-linear deformations when multiple

hand images are acquired from the same person.

That is demonstrated in Fig. 2 where a user’s thumb

appears to be longer in one of the images, a fact

which makes the system reject its true identity. This

problem is alleviated if the thumb (which can de-

form more than the other four ﬁngers) is excluded

from the feature vector calculation. Figure 1(b)

compares the ROC curves corresponding to using

all ﬁve ﬁngers versus excluding the thumb and/or

the little ﬁnger. The system which excludes the

thumb exhibits a substantially better performance

over the system which uses all ﬁve ﬁngers.

Summary

Hand shape-based biometric systems have been success-

fully demonstrated for applications involving personal

identity veriﬁcation. Their ease of use, nonintrusiveness,

public acceptance, integration capabilities, and small

resource requirements have made hand shape popular

among the different biometric modalities.

Related Entries

▶ Hand geometry

▶ Independent Component Analysis

▶ Multimodal systems

▶ Palmprint matching

References

1. Amayeh, G., Bebis, G., Erol, A., Nicolescu, M.: Peg-Free Hand

Shape Veriﬁcation Using High Order Zernike Moments. In:

Proceedings of the IEEE Workshop on Biometrics in conjunc-

tion with CVPR06, New York, June, 2006

2. Ross, A., Jain, A.K.: Biometrics. Encyclopedia of Biometrics

3. Yo

k, E., Dutagaci, H., Sankur, B.: Hand Biometrics. Image Vis.

Comput. 24, 483–497 (2006)

4. Jain, A.K., Duta, N.: Deformable Matching of Hand Shapes for

User Veriﬁcation. In: Proceedings of the IEEE International

Conference on Image Processing (ICIP), Kobe, Japan, 857–861

(1999)

5. Veldhuis, R.N.J., Bazen, A.M., Booij, W., Hendrikse, A.J.:

Hand-Geometry Recognition Based on Contour Parameters.

Hand Shape. Figure 3 Hand shape alignment. Two scans of different hands: (a–b) original images, (c) hand shapes

extracted from (a) and (b) overlaid, and (d) ﬁnger aligned shapes (Mean alignment error = 2.02 pixels).

686

Hand Shape

In: SPIE Biometric Technology for Human Identiﬁcation II,

Orlando, FL, 344–353 (2005)

6. Yo

k, E., Konukoglu, E., Sankur, B., Darbon, J.: Shape-Based

Hand Recognition. IEEE Trans. Image Process. 15(7), 1803–1815

(2006)

Hand Shape Biometrics

▶ Hand Geometry

Hand Silhouette Data

▶ Hand Data Interchange Format, Standardization

Hand Structure

▶ Anatomy of Hand

Hand Vascular Recognition

▶ Back-of-Hand Vascular Recognition

Hand Veins

GRA HAM LEEDH AM

School of Information and Communication Grifﬁth

University, Queensland, Australia

Synonyms

Finger vein; Palm dorsal vein; Palm vein

Definition

In the human hand there is a complex structure of veins

and blood vessels, many of which are just a few milli-

meters below the skin surface. Using noninvasive and

safe imaging techniques it is possible to capture an

image of the larger veins and blood vessels near the

skin surface in various parts of the hand. These images

are most readily obtained from the back of the hand and

the palm of the hand. This vein structure, which is

mostly invisible to the human eye, forms a pattern of

interconnecting lines which is different from one indi-

vidual to another and can be used as a physiological

biometric. Two imaging methods can be used for safe,

noninvasive imaging of veins near the skin surface:

(1) Far infrared thermography, and (2) near infrared

imaging. Far infrared imaging detects heat radiated

from the hand and veins. Near infrared imaging detects

infrared light reﬂected from a hand illuminated by near

infrared light.

Introduction

In recent years, vein pattern biometrics has attracted

increasing interest from research communities and

industry. A system that scanned the back of a clenched

ﬁst to determine hand vein structure for verifying user

identity was ﬁrst reported in 1991 by Cambridge Con-

sultants Ltd., in collaboration with the British Technol-

ogy Group (BTG), who had been studying the hand

vein pattern concept with the aim of developing a

commercial system which they call Veincheck [1].

Though their product did not achieve much commer-

cial success, the concept of hand vein patterns as a

biometric was founded and has recently attracted fur-

ther research and development interests to acquire the

vein patterns in the back of the hand [2–5] and in the

palm [6–8] as well as in the ﬁngers [9].

A vein pattern is the vast network of blood vessels

within a person’s body carr ying blood back to the

heart. Anatomically, the distribution of veins in the

body creates a vascular pattern which is believed to be

distinct from person to person [10] and is also ob-

served to be different between identical twin s. The

vascular patterns are reported to be stable over a long

period of time, as a person’s pattern of blood vessels

is ‘hardwired’ into the body at birth, and remains

relatively unaffected by aging, except for predictable

Hand Veins

687

growth as seen in ﬁng erprints. In addition, as the blood

vessels are hidden underneath the skin and are invisible

to the human eye, vein patterns are more difﬁcult to

copy or forge as compared to many other biometric

features.

The properties of probable uniqueness, stability,

and strong immunity to forgery of the vein pattern

make it a potentially good physiological biometric

to provide more secure and reliable person veriﬁcation.

However, as the vein patterns formed by superﬁcial

blood vessels lie underneath the skin surface, the invis-

ibility of veins to simple visual inspection system cre-

ates signiﬁc ant difﬁculties in the acquisition of the

vein pattern images. As the quality of the images

plays a key role in all the subsequent processing stages

of a vein pattern biometric system, the image acquisi-

tion is critical. In vein imaging for medical purposes,

X-rays and ultrasonic scanning are used to obtain

vascular images. While these methods can produce

high quality images of blood vessels, X-ray imaging

requires the invasive injection of a contrast agent into

the blood stream and dosage of ionizing radiation

which is dangerous with repeated exposure of even

low level radiation. Ultrasonic imaging, while not

known to have any adverse side effects, requires the

application of a gel to the skin to improve the trans-

mission of the sound waves as well as operator skill to

obtain a good image. These constraints are not accept-

able in general purpose biometric applications for se-

curity screening. Obtaining the vein pattern images

quickly and accurately in a nonintrusive and noninva-

sive manner is a key challenge in the vein pattern

biometric system.

Currently, the most effective means of obtaining

images of veins near the surface of the skin without

any invasive procedure or potentially dangerous side

effect is to use the infrared range of the

▶ electromag-

netic spectrum. Infrared imaging provides a contact-

less data acquisition method and requires no injection

of any agents into the blood vessels. In the electromag-

netic spectrum infrared refers to a speciﬁc region with

wavelength typically spanning from 0.75 to 1,000 mm.

This region is commonly further divided into four

sub-bands: (1) Near infrared (0.75–2 mm); (2) Middle

infrared (2–6 mm); (3) Far infrared (6–14 mm);

(4) Extreme infrared (14–1,000 mm). Imaging objects

within these four regions operates using different phys-

ical mechanisms and it results in imag es with signi-

ﬁcantly different properties. Far infrared and near

infrared are the most suitable to capture images of

human bodies.

Far Infrared Imaging

All objects emit infrared radiation when they are

heated. The far infrared imaging technology forms an

image passively using the infrared radiation emitted

by the human bod y.

Principle of Far Infrared Imaging

The total emissive power w is described by the Stefan-

Boltzmann Law given in (1), where x is the emissivity

of the object and s =5:6703  10

8

watt/m

Stefan’s constant. The relationship between the wave-

length l and black body temperature T is formulated

by Wiens Displacement Law bas ed on Planck’s energy

distribution law given in (2).

w ¼ x  s  T

ð1Þ

max

¼ 2 : 9  10

3

=T ð2Þ

Typically, a human body emits infrared radiation with

wavelength in a range of 3–14 mm. These infrared waves

radiate into the atmosphere and are attenuated accord-

ing to the infrared transmittance spectrum of the atmo-

sphere. At wavelengths of 3–5 mmand8–14mm, the

radiant emittance of the infrared spectrum possesses

the highest transmittance rate. Therefore, by using a

thermal camera with detector sensitivity in the range of

either 3–5 mmorof8–14mm, an image showing the heat

distribution of the human body can be obtained. Medi-

cal researchers have observed that superﬁcial veins have

slightly higher temperature than the surrounding tissue.

Therefore, using thermal imaging, an image of the heat

distribution of body will display the location of veins

just below the surface of the skin.

To acquire a far infrared image of the hand, the

hand is usually placed on a ﬂat surface with a far

infrared camera focused on the hand from above.

The far infrared camera will capture the temperature

proﬁle of the hand and transfer it to a computer

which can convert the temperature data into either

grey scale or color coding for display on a standard

visual display and stored as a digital image for later

computer processing.

688

Hand Veins

Far-Infrared Vein Image Quality

Figures 1 and 2 show typical vein pattern images cap-

tured using a far infrared imaging method and con-

verted to grey scale images. The darker the grey scale

the cooler the pixel. The major vascular network in the

back of the hand is successfully captured and appear s

as light grey lines as shown in Fig. 2. The images in

Fig. 1 were captured in a normal ofﬁce environment

(approximately 20



C and 50% humidity) such that

there is sufﬁcient temperature difference at the skin

surface to distinguish the location of the major veins

beneath the skin. Figure 2 shows two images captured

outdoors in a hot tropical climate (30–34



C and >80%

humidity). In this case the temperature of the hand

is closer to the blood temperature and there is insufﬁ-

cient difference in radiated thermal energy where the

veins are located and the surrounding tiss ue to be

discernable using the 0.08



C resolution of the camera

used to capture this image. Far infra red imaging

technolog y is very sensitive to external conditions

which affect the temperature of the hand. In addi-

tion, far infrared imaging can only capture the major

vein patterns. The smaller capillaries are not visible

and the information contained in the large vein pattern

is limited.

Hand Veins. Figure 1 Examples of far infrared images of hands at room temperature mapped to grey scale.

Hand Veins. Figure 2 Examples of far infrared images of hands in a tropical climate mapped to grey scale.

Hand Veins

689