IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012 1511
The OU-ISIR Gait Database Comprising the Large
Population Dataset and Performance Evaluation o f
Gait Recognition
Haruyuki Iwama, Mayu Okumura, Yasushi Makihara, and Yasushi Yagi, Member, IEEE
Abstract—This paper describes the world’s la
rgest gait
database—the “OU-ISIR Gait Database, Large Population
Dataset”—and its application to a statistically reliable perfor-
mance evaluation of vision-based gai
t recognition. Whereas
existing gait databases include at most 185 subjects, we construct
a larger gait database that includes 4007 subjects (2135 males and
1872 females) with ages ranging fro
m 1 to 94 years. The dataset
allows us to determine statistically significant performance dif-
ferences between currently proposed gait features. In addition,
the dependences of gait-reco
gnition performance on gender and
age group are investigated and the results provide several novel
insights, such as the gradual change in recognition performance
with human growth.
Index Terms—Gait database, gait recognition, large population,
performance evaluation.
I. INTRODUCTIO N
B
IOMETRIC-BASED person-recognition techniques have
become increasin gly important in crime prevention and
investigation. Gait, as a biometric cue, has relatively r ecently at-
tracted much attention and it is expected to be applied to wide-
area su rveillance and crime investigation owing to the possi-
bility of identifying subjects from a distance without the coop-
eration of the subjects. Thus, vision-based gait-recognition ap-
proaches have been widely deve loped in recent years [1]–[9].
For the development and statistically reliable evaluation of
gait-recognition approaches, the construction of a common
gait database is essential. There are tw o considerations in
constructing a g ait database: (1) the variation in walk ing con-
ditions (e.g., v iew, speed, clothing, and carrying conditio ns),
and (2) the number and diversity of the subjects. The first
Manuscript received December 12, 2011; revised May 02, 2012; accepted
May 14, 2012. Date of publication June 11, 2012; date of current version
September 07, 2012. This work was supported in part by Grant-in-Aid for
Scientific Research (S) 21220003 and the “R&D Program for Implementation
of Anti-Crime and Anti-Terrorism Technologies for a Safe and Secure Society,”
Strategic Funds for the Promotion of Science and Technology of the Ministry of
Education, Culture, Sports, Scie nce and Technology, the Japanese Government.
The associate editor coordinating the review of this manuscript and approving
it for publication was Dr. Fabio Sco tti.
The authors are with the Department of Intelligent Media, Insti-
tute of Scientific and Industrial Research, Osaka University, Ibaraki,
Osaka, 567-0047, Japan (e-mail: iwama@am.sanken.osaka-u.ac.jp; oku-
mura@am.sanken.osaka-u.ac.jp; makihara@am.sanken. osaka-u.ac.jp;
yagi@am.sanken.osaka-u.ac.jp).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Di
gital Object Identifier 10.1109/TIFS.2012.2204253
consideration is important in evaluating the
robustness of
the gait recognition, because walking c
onditions d e pend on
the time and circumstances and often di
ffer between gallery
and probe. For instance, the clothing
and carrying conditions
when walking along a street in a suit w
ith a bag while o n
business can differ from those when
strolling empty-handed in
casual clothes during leisure tim
e. The second consideration
is important to ensure statist
ical re liability of the performance
evaluation. Moreover, if the
database is used for soft biometr ic
applications such as gait-b
ased gender and age classification
[10], [11], the diversity o
f subjects in terms o f gender and age
plays a significant r ole in
the performance evaluation.
Although several gait data
bases have been constructed
[12]–[25], with most of th
ese taking good account of the first
consideration, the sec
ond consideratio n is still insufficiently
addressed since these
databases i nclu de at most 18 5 subjects
[24] and the subjects
’ genders and ages are biased in many
of the d atabases. Th
e exceptions are the large-scale datasets
introduced in [26] a
nd [27], which do address the second con-
sideration and inc
lude respectively, 1,035 and 1,728 subjects
with ages rangin
g from 2 to 94 years. In these datasets, how-
ever, the gait i
mages are captured u sing cameras with varying
poses (e.g., a c
amera’s pose on one day differs slightly from
that on anothe
r day, or some subjects are captured using first
one camera an
d th en another with a slightly different pose) and
this coul d in
troduce bias into the evaluation results.
In this paper
, we focus on the second consideratio n and in-
troduce a la
rge population dataset that is a major upgrade to
previousl
y reported large-scale datasets in [26] and [27]. The
extension
s o f this dataset are as follows.
1) The number
of subjects is c onsiderably greater in the
dataset; i
.e., there are more than thrice the number of
subjects
in the d ataset in [26] and more than twice the
number i
n the d ataset in [27].
2) A ll s ilho
uette images are normalized with respect to the
image pl
ane to remove the bias of camera rotation f or more
equita
ble performance evaluation.
3) The obse
rvation angle of subjects in each fram e is specif-
ically
defined for the sake of fair analysis in terms of the
observ
ation a ngle, whereas previous works merely defined
the an
gle as a side view.
Our da
taset is the largest gait dataset in the world, comprisin g
over 4
,000 sub jects of both genders and includin g a wide range
of ag
es. Although the dataset does not include any variation s in
walk
ing conditions, it allows us to investigate the u pper limit of
gai
t-recognition p er formance in a more stat isti cally reliable way
1556-6013/$31.00 © 2012 IEEE
1512 IEEE T R ANSACTIONS ON INFORM ATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012
TABLE I
E
XISTING MAJOR GAIT DATABASES
and to reveal how gait-recognition performance differs between
genders and age groups. Thus, our dataset can contribute much
to the develop men t of gait-based applications, and we demon-
strate its validity through experiments with state-of-the-art gait
representations.
The outline of the paper is as follows. Section II introduces
existing gait databases, while Section III addresses the construc-
tion of the dataset. The gait-recognition approach for perfor-
mance e valu ation is descr ibed in Section IV, and various perfor-
mance evaluations usin g our dataset are presented in Sectio n V.
Section VI p resents our conclusions and discusses future work.
II. R
ELATED WORK
Existing major gait databases are summarized in Table I.
Here, we briefly describe these databases.
The Soton database is compo sed of a sm all population d ataset
[15] and a large population dataset [14]. T he small dataset con-
tains subjects walking around an indoor track, with each sub-
ject filmed wearing a variety of footwear and clothing, carrying
various bags, and walking at different speeds. Hence, the data-
base is used for exploratory factor analysis of gait recognition
[29]. The large dataset was the first gait database to contain over
100 subjects and h as contributed to the study of gait recognition
mainly in terms of intersubject variation. The recently published
Soton Temporal database [21] contains the largest time v aria-
tions; up to 12 months to date [28]. It e nabl es the investigatio n
of the effect of time on the performance of gait biometrics, al-
lowing the use of 3-D volumetric data.
The USF dataset [18 ] is one of the most widely used gait
datasets and is composed of a g allery and 12 probe sequences
captured outdoors under different walking co nditions including
factors such as view, shoes, surface, baggage, and time. As the
number of factors is the largest of all existing databases, despite
there being only two variations for each factor, the USF d atabase
is suitable for the evaluation of the interfactor effect, as opposed
to the intrafactor effect, on gait-recognition performance.
The CASIA database, Dataset A [16] contains image se-
quences from three views and can be used for the analysis
of the effect of the v iew angle on recognition performance.
The CASIA database, Dataset B [19] consists of m ultiview
(11 view s) walking sequences and includes variations in the
view angle, clothing, and carrying conditions. Since it contains
the finest azimuth view variations, it is useful for the analysis
and modeling of the effect of view on gait recognition [30].
The CASIA database, Dataset C [20] was the first database to
include in frared gait images captured at night, thus enabling
the study of night gait r ecogn ition.
The OU-ISIR Gait Database, Treadmill Dataset [22]–[25]
contains gait images of subjects on a treadmill with the largest
range of view variations (25 views: 12 azimuth view s times 2
tilt angles, plus 1 top view), speed variations (9 speeds: 1 km/h
intervals between 2 and 10 km/h), and clothing variations (up
to 32 combinations), and as such, it can be used to evaluate
view-invariant [4], speed-invariant [22] and clothing-invariant
[23] gait recognition. In add ition, it is used to analyze gait
features in gender and/or age-group classification [25], since
the diversities of gender and age of the subjects are greater than
those in currently available gait databases.
Next, we review the number and diversity of subjects. Table I
shows that existing major databases includ e more than 100
subjects. Although these databases are statistically reliable to
some extent, the number of subjects is insufficient when com-
pared w ith databases of other biometrics such as fingerprints
and faces. In addition, the populations of genders and ages are
biased in m any of these databases; e.g., there are no children
in the USF dataset with m ost of the subjects in their twenties
and thirties, while the ratio of males to females is 3 to 1 in
the CASIA dataset (Dataset B). Such biases are undesirable in
performance evaluation of gait-based gender and age-group
estimation and in performance comparison of gait recognition
between genders and age groups.
III. T
HE OU-ISIR GAIT DATABASE,LARGE POPULATION
DATAS ET
A. Capture System
An overview of our capture system is illustrated in F ig. 1.
Each subject was asked to walk at his or her own preferred speed
through a straight course (red arrows) at most twice u nder the
same conditio ns. The length of the course was approximately
10 m, with approximately 3 m (at least 2 m) sections at the
beginning and end regarded as acceleration and deceleration
IWAMA et al.: OU-ISIR GAIT DATABASE COMPRISING THE LA RGE POPULATION DATASET AND PERFORMANCE EVALUA
TION 1513
Fig. 1. Overview of capture system and captured images.
zones, respectively. Two cameras were set approximately 4 m
from the w alkin g course to observe (1) the transition from a
front-oblique view to a side view ( camera 1), and (2) the transi-
tion from a side view to a rear-oblique view (camera 2). We used
Flea2 cameras manufactured by Point Gray Research Inc. with
HF3.5M-2 lenses manu factu red by SPACE Inc. T he image size
and frame rate w ere, respectively, 640
480 pixels and 30 fps.
The recorded image format was uncompressed bitmap. More-
over, green background panels and carpet (if available) were
arranged along the walking course for the purpose of clear sil-
houette extraction.
B. Data Collection
The dataset was collected during entertainment-oriented
demonstrations of an online gait personality measurement
[31] at outreach activity events in Japan, including the Dive
Into the Movie project (DIM2009) [32], the 5th Regional
Disaster and C rim e Prevention Expo (RDCPE2010), Op e n
Campus at Osaka University (OU -O C 201 0/2 011), and the
Core Research for Evolutional Science and Technology
project (http://www.jst.go.jp/kisoken/crest/en/index.html,
CREST2011). All the events were held at indoor h alls and the
numbers of visitors at each event are summarized in Table II.
Each subject was requested to give their informed consent
permitting the use of the collected data for research purposes.
Also, the age and gender of each subject were collected as meta-
data. All the subjects walked empty-handed, wearing their own
clothing (some subjects wore a hat) and footwear. Examples o f
images captured at each event are shown in Fig. 1 .
C. Statistics
From the data collected by camera 1 (im ages were taken
with two cameras at the events), the world’s largest gait dataset
of 4,007 subjects (2,135 males and 1,872 females) with ages
ranging from 1 to 94 years was constructed. We call this d ataset
the “OU-ISIR Gait Database, Large Population Dataset C1
Version1”
1
, which we abbreviate to OULP-C1V1
2
. Detailed
distributions of the subjects’ gender and age are shown in
Fig. 2, w hil e example images of the sub jects are shown in
Fig. 3. A lmost all the subjects are of Asian descent.
The dataset comprises two subsets, which we call
OULP-C1V1-A and OULP-C1V1-B. OULP-C1V1-A is
a set of two sequences (gallery and probe sequences) per
subject and is intended for use in evaluating gait-recogni-
tion performance under almost con stant walking conditions.
OULP-C1V1-B is a set of one sequence per subject and is
intended for use in investigating gai t-b a sed gender classifica-
tion and age estimation. OU LP-C1V 1-A and OULP-C1V1-B
are major upgrades to the datasets introduced in [26] and [27],
respectively. For brevity, we omit the descr iption of the dataset
header “OULP-C1V1-”.
Each of the main subsets is further divided into five subsets
based on the observation angle (55 [deg], 65 [deg], 75 [deg],
85 [deg], and including all four angles) of each subject. We call
these subsets A/B-55, A/B-65, A/B-75, A/B-85,andA/B-ALL,
respectively, with each subject belonging to at least one of these
subsets. The observation angle
of each subject in each frame
is defined by the y-axis of the world coo rdinate system (which
is parallel to the walking direction) and the line of sight of the
camera as illustrated in Fig. 4.
A subject is included in a bin of a subset if one gait period oc-
curs in the ran ge of angles (as illustrated in Fig. 4) corresponding
to that subset. For example, if a subject is recorded twice ( both
gallery and probe sequences) w ith a complete g ait period in the
range of 55 [deg], the subject is included in a bin of A-55 and
one of B-55. Moreover, if a subject is recorded twice with a
complete gait period covering all the angle ranges, the subject
is included in a bin of all the subsets. A gait period is calculated
from th e whole sequence (see Section IV-B for details on the
calculation of the gait period).
An example image for each observation angle is shown in
Fig. 4, while a breakdown of the number of subjects is giv en in
Table III. In this tab le, the values in the “Total” colu mn re pr esen t
the number of subjects included in at least one of the subsets of
55 [deg], 65 [deg], 75 [deg], an d 85 [deg]. As mentioned above,
the numbers of subjects for dataset A represent those that have
been recorded twice. Also, the differences between datasets A
and B for each subset represent the numbers of subjects recorded
1
To be prepared for publication. The data w ill be published in the form of
normalized silhouette image sequences in PNG format, with a total data size of
about 1.5 G B.
2
The naming for mat is OULP-[camera ID][version ID]-[header1]-
[header2]-….
1514 IEEE T R ANSACTIONS ON INFORM ATION FORENSICS AND SECURITY, VOL. 7, NO. 5, OCTOBER 2012
TABLE II
V
ISITORS AT EVENTS
Fig. 2. Distributions of the subjects’ gender and age in OULP-C 1V1.
Fig. 3. Examples of subjects in OULP-C1V1.
only once. Take for example, the subset of 55 [deg] in Table III
(A-55 and B-55) where 3,706 subjects are recorded twice and
292 subjects are recorded only once. Note that there are also
differences in the n um bers of subjects between subsets, because
the sequence length and observ ation angles for each subject are
not exactly the same.
D. Advantages
Compared with existing gait databases, our dataset has the
following strengths.
1) Large population: The numb er of subjects is m ore than 20
times that in publicly available large-scale gait databases.
This impr oves the statistical r eliability of variou s perfor-
mance evaluations such as the comparison of gait recogni-
tion.
2) Gender balance: T he ratio of males to females is close
to 1. This is a desirable property for more reliable perfor-
mance evaluation of gait-based gender classification and
for comparison of gait-recognition perfo rm ance between
genders.
3) W hole generation: The age range is from 1 to 94 y ears
with each 10-year interval up to 49 years of age con-
taining more t han 400 subjects (even in the smallest
Fig. 4. Definitions of the world coordinate system and the o bser v ation angle
of a subject, and an exam ple image at each observation angle. The
plane
corresponds to the background wall behind the walking subjects, while the
plane corresponds to the ground plane.
subset A-ALL). In addition, it is noteworthy that our
dataset includes a sufficient number of children at all
stages of growth, whereas other large-scale gait databases
are mainly co mposed of adult subjects. This provides
more statistically reliable results for gait-based age-group
classification and comparisons of the difficulties in gait
recognition among age groups.
4) Silhouette quality: The quality of each silhouette image
is relatively high because we visually checked each sil-
houette more than twice and made manual modifications
if necessary. This enables the elimination of s ilhouette
quality problems from g ait analysis. On the con trar y, the
silhouette images in most of the existing public databases
are automatically extracted and often include significant
over/under-segmentation. Although manually modified
silhouettes were created in the investigation of the effect
of silhouette quality on gait recognition in [33] and [34],
these have not been published.
E. Preprocessing
This section briefly describes the method used for size-nor-
malized silhouette extractio n.
1) Silho uette Extraction: The first step involved extraction
of gait silhouette images via graph-cut-based segmentation [35]
in conjunction with background subtraction. Of course, over/
under-segmentation e rrors appeared in some extracted silhou-
ette images. Hence, as described above, we visually checked
all silhouette images at least twice and then manually m odi-
fied under/over-segmentation if necessary. In more detail, a sil-
houette was shown to the observer in the form of a com po site
image in wh ich the silhouette contour was overlaid on th e cor-
responding original image. The observer checked whether the
silhouette contour fitted the visually p erceived human contour
and if not, modified it using a GUI tool specially developed for
this pu rpose.
2) Correction of Camera Rotation: In the second step,
image normalization, in cluding the co rr ection of distor tion and
camera rotation, was carried out. Because the camera pose in
the world coordinate system for each day/event was not strictly
the same, we normalized the camera rotations in all silhouette
images such that the i mage plane in each is parallel with the
plane in the world coordinate system as show n in Fig. 5.
IWAMA et al.: OU-ISIR GAIT DATABASE COMPRISING THE LA RGE POPULATION DATASET AND PERFORMANCE EVALUA
TION 1515
TABLE III
B
REAKDOWN OF THE NUMBER OF SUBJECTS IN OULP- C1V1 (MALE/FEMALE)
Fig. 5. Examples of th e original and no rmalized camera pose, image plane, and
images. In the rotation-corrected image in (b), the set of cyan lines and set of
magenta lines represent the sets of parallel lines in the scene used to determine
the vanishing points, while the white dashed lin e rep resents the vertical center
line of the image. The obs er vation angle is 90 [deg] at this line. (a) Original
image plane. (b) Normalized image plane.
First, the intrinsic parameters of the camera and coefficients of
lens distortion were estim a ted [36]
3
and distortion corrected.
An example of an u ndistorted image is shown i n Fig. 5(a). The
transformation m a tr ix of cam er a rotat ion f rom the original pose
(shown in Fig. 5(a)) to the target pose (shown in Fig. 5 (b)) was
then estimated for each day/event from the undistorted im age
using a pair of vanishing po ints [37] (i.e., horizontal an d ver-
tical vanishing points), estimated from the sets of parallel lines
in the scene [38]. Finally, all the image pixels in t he original
image plane were reprojected onto the normalized image plane.
An exam ple of a camera rotation corrected image is shown
in Fig. 5(b). Also, examples of a subject in each dataset after
rotation correction are shown in Fig. 6.
3) Registration and Size Normalization: The third step in-
volved registration and size normalization of th e silhouette im-
ages [4]. First, the top, bottom , and horizontal center of the sil-
houette regions were obtained for each fram e. The horizontal
3
Calibration procedures were implemented using OpenCV version 1.1 fun c-
tions.
center was chosen as the median of the horizontal positions be-
longing to the region. Second, a moving -av erage filter was ap-
plied to these positions. Third, we n ormalized the size of the
silhouette images such that the height was just 128 pixels ac-
cording to the average positions, and the aspect ratio of each
region was maintain ed. Finally, we produced an 88
128 pixel
image in which the average horizon tal m edian c orresp ond s t o
the horizontal center of the image. Exam ples of size-normal-
ized silhouettes are shown in Fig. 7.
IV. G
AIT RECOGNITION
This section describes a framework for performance evalua-
tion of gait recognition.
A. Gait Features
The current trend in gait representation is appearance and pe-
riod-based representation, such a s the averaged silhouette [39],
also known as the G ait Energy Image (G EI) [ 40]. In this paper,
we deal with six such state-of-the-art gait features: GEI, Fre-
quency-Domain Feature [4] (referred to as FDF in this paper),
Gait Entropy Image (GEnI) [5], Masked GEI based o n GEnI [7]
(referredtoasMGEIinthispaper), Chrono-Gait Image (CGI)
[6], and Gait Flow Image (GFI) [8].
The GEI is obtained by averaging silhouettes over a gait
cycle, while the FDF is g ener ated by applying a Discrete
Fourier Transform of the temporal axis to the silhouette im ages
in a gait cycle. In this paper, 0, 1, and 2 tim es frequency ele-
ments are used. The GEnI is computed by calculating Shannon
entropy for every pixel over a gait cycle, where the value of
the GEI is regarded as the probabili ty that the pixel takes the
binary value. The MGEI is computed by masking the GEI with
a pair-wise mask generated by each pair of probe and gallery
GEnIs. The GEnI and MGEI aim t o select the dynamic area
from the GE I. The CGI is a tem poral template in which the
temporal information among gait frames is encoded by a color
mapping function, and is obtained by compositing the color
encoded gait contour images in a gait cycle. The GFI is based
on an optical flow field from silhouettes representing motion
information and is created by averaging the binarized flow
images over a gait cycle. An example of each feature is show n
in Fig. 8.
B. Gait Period Detection
For the quantification of periodic gait motion, we adopted
the Normalized Auto Correlation (NAC) of the size-normalized
