International Journal of Machine Learning and Computing, Vol. 1, No.

4, October 2011

Biometrics System based on Human Gait Patterns

Praveen Gupta, Rajendra Singh, Rohit Katiyar, Member, IACSIT and Ritesh Rastogi

the length of steps. But what seems to be an easy task for

Abstract— Today’s commercially available biometric humans must not necessarily apply to computers.
systems show good reliability. However, they generally lack Although those early results were encouraging and
user acceptance. In general, people favour systems with the promising, gait has not been proposed as a biometric feature
least amount of interaction. Using gait as a biometric feature
until recently. Possible reasons might encompass the lack of
would lessen such problems since it requires no subject
interaction other than walking by. Consequently, this would reliable and inexpensive sensors as well as the lack of
increase user acceptance. And since highly motivated users processing power to handle the huge amount of data.
achieve higher recognition scores, it increases the overall All of the aforementioned methods and approaches can be
recognition rate as well. The latest research on gait-based roughly divided into two groups. Namely, the model-free
identification—identification by observation of a person’s and the model-based approaches. Model-free approaches
walking style provides evidence that such a system is realistic
have no underlying three-dimensional representation of a
and is likely to be developed and used in the years to come.
This article outlines the application of gait technologies for walking person and mainly rely on statistical properties of
security and other purposes. Gait analysis and recognition can the acquired gait data. Conversely, the model-based methods
form the basis of unobtrusive technologies for the detection of have a model of the human body, or at least part of it, that is
individuals who represent a security threat or behave fit to every frame of the walking sequence. In order to fit the
suspiciously. model in the frame, static parameters such as the limb
lengths, body height, body width as well as dynamic
Index Terms— Gait Recognition System, Holistic Approach,
Model Based Approaches, Pattern Recognition
parameters such as the angular velocities and walking speed
need to be estimated. Research conducted thus far in the
area of gait recognition has shown that gait can be reliable in
I. INTRODUCTION combination with other biometrics. If we assume that palm,
fingerprint, and iris methods belong to a different (obtrusive)
Gait is not a new topic in research and scientific literature.
class of biometrics, additional biometrics that could be used
It has been investigated and examined in various aspects
in conjunction with gait in a multibiometric system would
over the past decades. On the one hand, research was
be face, ear and foot pressure [2]. In a multibiometric
inspired by medical applications to track rehabilitation or as
system, gait and foot pressure could be used to narrow down
a diagnostic tool. On the other hand, research was also
the database of subjects. Subsequently, face recognition
driven by the sport shoe industry. Murray conducted in 1967
could be used for identification of a test subject among the
[37] a systematic study to fully characterize the coordinated
reduced set of candidate subjects. Otherwise, the three
movement patterns of the various parts of the body that
biometrics could be combined altogether, e.g., using the
constitute the walking act. His empirical investigation was
techniques described in [3]. The combination of gait with
based on a relatively large sample set of 60 normal men in
face recognition was examined in [4] and [5]. In [5], it was
wide ranges of age and height. He obtained the walking
shown that gait is more efficiently utilized in a multimodal
patterns with reflective targets attached to specific
framework when it is combined directly with the facial
anatomical landmarks which he illuminated with a strobe-
features rather than preceding the face recognition module
light flashing 20 times per second. The study suggests that
as a filter.
gait is a unique personal characteristic, if all gait movements
are considered; this indicates that gait could be used as a
II. TERMOLOGY
promising feature for biometric authentication.
Later, in 1977, Cutting and Kozlowski [38] empirically Despite the differences among walking styles, the process
showed that recognizing friends by their gait is indeed a of walking is similar for all humans. A typical sequence of
surprisingly simple task for humans; even when stripped stances in a gait cycle is shown in Figure 1. A detailed
from all familiarity cues such as clothing and hairstyle. analysis of gait phases can be found in [6]. For simplicity,
Light sources mounted on joints that are prominent during we consider the following four main walking stances [7]:
the act of walking were sufficient for identification. It is right double support (both legs touch the ground, right leg in
noteworthy that people recognized others not by using static front), right midstance (legs are closest together, right leg
properties such as height but dynamic aspects such as touches the ground), left double support, and left midstance.
amount of arm swing, rhythm of the walker, bounciness, or Although some other definitions would also be appropriate,
in this article we define a gait cycle as the interval between
two consecutive left/right midstances. The interval between
Manuscript received July 13, 2011, revised September 30, 2011.
any two consecutive midstances is termed half cycle. The
Praveen Gupta, Rajendra Singh and Rohit Katiyar are with Dept. CSE,
PSIT, Kanpur, India. (e-mail: [email protected]; time interval in which a gait cycle is carried out is called the
[email protected]; [email protected]) gait period, whereas the walking frequency is termed the
Ritesh Rastogi is with MCA Dept., NIET, Noida, India. (e-mail: fundamental gait frequency.
[email protected])

378
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

III. A GENERIC GAIT RECOGNITION SYSTEM particular subjects and varies between subjects and can thus
Gait recognition is a multistage process (see Fig 2). It is be used as a biometric feature.
important that gait capturing is performed in environments Huang et al. suggested two different approaches in their
where the background is as uniform as possible. Moreover, publications using characteristics extracted from video
since gait recognition algorithms are not, in general, sequences. The first approach is based on spatial templates
invariant to the capturing viewpoint, care must be taken to [41] of the subject’s binarized silhouette, whereas the
conduct capturing from an appropriate viewpoint. Preferably, second uses temporal-templates [41-42, 44-45] of the
the walking subject should be walking in a direction silhouette. In both cases a combination of an Eigenspace
perpendicular to the optical axis of the capturing device Transformation (EST) and Canonical Space Transformation
since the side view of walking individuals discloses the most (CST) [43] are applied to reduce data dimensionality and to
information about their gait. Once a walking sequence is circumvent the singularity problem that occurs in the CST,
captured, the walking subject is separated from its when the number of elements in the feature vector is higher,
background using a process called background subtraction. than the number of feature vectors in the training set.
A critical step in gait recognition is feature extraction, i.e., In [46], Nash et al. proposed a new model-based
the extraction, from video sequences depicting walking technique to allow the automated determination of human
persons, of signals that can be used for recognition. This gait characteristics. Their technique employs a parametric
step is very important since there are numerous conceivable two-line model representing the lower limbs. To speed up
ways to extract signals from a gait video sequence, e.g., the search of the parameter space, they used a genetic
spatial, temporal, spatiotemporal, and frequency-domain algorithm (GA) based implementation of the Velocity
feature extraction. Therefore, one must ensure that the Hough Transform (VHT) rather than an exhaustive search.
feature extraction process compacts as much discriminatory Although their approach is promising, the accuracy of the
information as possible. Finally, there is a recognition step, estimated hip rotation patterns is still insufficient for
which aims to compare the extracted gait signals with gait biometric purposes.
signals that are stored in a database. Apart from the apparent Meyer et al. described in [47] a system based on statistical
usefulness of gait analysis in biometric applications, gait has models that performs automatic classification of different
several important nonbiometric applications. gaits from grey-level image sequences. In particular, they
can differentiate between walking, running, hopping, and
limping. To extract the trajectories of the different body
parts they used statistical models. The classification is
performed with discrete Hidden Markov Models (HMM).
A different approach was followed by Orr and Bawd [48]
who proposed a method using simple parameters extracted
from the ground reaction force profiles (GRF). To
Fig. 1. Different silhouette stances during a gait cycle. characterize each footstep profile, they propose ten features
(mean value of the profile, its standard deviation, length of
the profile, area under the profile, x-y-coordinates of the two
IV. SUMMARY ON PREVIOUS WORK maximum points and the minimum point). The poor
The study of gait as a discriminating trait was first recognition rate of this simple approach limits its
attempted a few decades ago from a medical/behavioral applicability for low-security environments only. However,
viewpoint [8], [9]. Murase and Sakai developed [Murase96] the method is perfectly suitable for its intended purpose in
a method to efficiently calculate the spatio-temporal the Aware Home Research Initiative (AHRI).
correlation for model-free moving object recognition. To In September 2000, the DARPA launched the HumanID
lower the computational cost of the spatio-temporal program with 26 individual projects and research groups
correlation they reduced the dimension of the input vectors involved from the USA, Germany, and England. The goals
with an orthogonal transformation and performed the of the project are to develop non cooperative, multimodal
correlation in the resulting low-dimensional parametric surveillance technologies for identifying humans at a
eigenspace representation. This general approach can be distance under day/night, and all-weather conditions. The
applied not only to gait but to other moving object HumanID program has two phases: The initial 2 years of
recognition problems as well. Phase I will end in late 2002 with a major evaluation. Phase
In 1997 Addlesse et al. proposed in [39] an Active Floor II lasts another 2 years and continues research with the most
system. They used an array of four by four load cells to promising approaches identified in the technology
measure the force, perpendicular to the floor, exerted by a assessment at the end of Phase I.
walking person. To characterize the footsteps a Hidden Although most of the research projects are still in an early
Markov Model (HMM) was trained using data acquired stage, some groups have already published preliminary
from 15 different individuals. The best HMM-configuration results.
achieved a recognition rate of 91 %. Recently, Bobick and Johnson published two papers
In [40], Little and Boyd theorized an alternate video [49,50] where they proposed a multi-view method that
based method. Their description of the spatial distribution of recovers body and stride parameters of the subjects as they
optical flow yields model-free frequency and phase features walk. In particular they estimate four static distances: the
whose variation over time is periodic. The relative phase vertical distance between the head an the foot, the distance
difference among these periodic signals is repeatable for between the head and the pelvis, the maximum distance

379
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

between the foot and the pelvis, and finally the maximum frames, of a limited number of representative frames for
distance between the left and right foot during the double each sequence. This process seems to capture all structural
support phase. Instead of reporting percent of correct information in a gait sequence while implicitly yielding
matches from a limited database (20 subjects), they denoised frames that can be used directly for recognition.
introduced a novel confusion metric that allowed them to This approach is taken in [23] and [24]. In [23], the
predict how their static body parameters discriminate even recognition is based on the comparison of such templates,
in a large population. whereas in [24], the templates are derived in the context of
Gene Greneker’s group at the Georgia Tech Research training an exemplar- based hidden Markov model (HMM)
Institute is working on a radar device that can be used to that additionally takes into account the gait dynamics. All
record the human gait signature over a distance of up to 120 methods in this class yield state-of-the-art performance.
meters.
J. Shi’s research group at the Carnegie Mellon University V. GAIT CYCLE DETECTION
has already published a technical report [51] detailing the For the study of gait analysis, we assume that the walking
capturing of 25 individuals walking on a treadmill in the subject has been extracted from a gait sequence using
CMU 3D room. The subjects performed four different standard image processing techniques. An important part of
activities: slow walk, fast walk, inclined walking, and the gait analysis process is gait cycle detection, i.e., the
walking with a ball while being filmed using six color partitioning of a gait sequence into cycles that depict a
cameras with different viewing angels. complete walking period. In [13], although no explicit cycle
J. Phillips et al. from the NIST14 will publish a proposal partitioning was attempted, a method using linear prediction
[52] of a reference implementation of a biometric system was proposed for fitting a sinusoidal signal to the noisy
using gait analysis. This baseline algorithm will be used to extracted signals. In [24], an adaptive filter was used to filter
characterize the conditions under which the problem of the foreground sum signal prior to the calculation of the gait
identifying/authenticating people using gait is solvable Later, cycles using the minima of this signal. In [25], the
several attempts were made to investigate the gait autocorrelation of the foreground sum signal was taken to
recognition problem from the perspective of capturing and calculate the walking period and compute the coefficients of
analyzing gait signals [10]–[14]. The techniques used for an optimal filter for the denoising of the sum signal.
gait recognition can be divided into two categories: holistic
(feature/appearance based) and model based. Techniques
that address the gait recognition problem using only
sequences of binary maps of walking human silhouettes are
of much interest since they do not presume the availability
of any further information, such as color or grey-scale
information, which may not be available or extractable in
practical cases. The main focus of algorithms is the tracking Fig. 2. Block Diagram of a Gait recognition/authentication system.
of silhouettes, analysis of the tracked silhouettes for feature
extraction purposes, and recognition using the extracted VI. MODEL BASED APPROACHES
features. A baseline method was proposed by the University Model-based approaches employ models whose
of South Florida [15]. It was tested on a gait database that is parameters are determined using processing of gait
tailored to the study of the impact of several factors, such as sequences [22], [26], [27], [28]. Unlike holistic approaches,
viewpoint, footwear, and surface, on the performance of a model-based approaches are, in general, view and scale
gait recognition algorithm. invariant. This is a significant advantage over the holistic
The deployment of motion fields in gait recognition was approaches since it is highly unlikely that a test gait
investigated in [13] and [16]. Although both methods sequence and a reference sequence will be captured from
reported good results on their own databases, they presume identical viewpoints. However, since model-based
the availability of texture information, which must be used approaches rely on the identification of specific gait
for the accurate computation of the motion fields. In [17]– parameters in the gait sequence, these approaches usually
[20], several feature extraction techniques were proposed require high-quality gait sequences to be useful.
based on the calculation of projections, contours, or other A multiview gait recognition method was proposed in [26]
such features from gait silhouettes. A comparison among using recovered static body parameters, which are
different features will be presented later in this article. In measurements taken from static gait frames. Gait dynamics
[21], a comparison is provided of several techniques for are not used. The static parameters used in [26] are the
improving the quality of silhouettes extracted from video height, the distance between head and pelvis, the maximum
sequences depicting humans walking. The silhouettes were distance between pelvis and feet, and the distance between
extracted using a model-based method that produces the feet [Figure 3(a)]. The static parameters are view
silhouettes that have fewer noise pixels and missing parts. invariant, which makes them very appropriate for
The resulting sequences were tested with the model-based recognition applications. In [22], the silhouette of a walking
algorithm in [22], and the overall system was shown to person was divided into seven regions. Ellipses were fit to
improve on the baseline system in [15]. To the authors’ each region [Figure 3(b)] and region feature vectors were
judgment, the most promising approach for gait recognition formed, including averages of the centroid, the aspect ratio,
is based on the formation, by means of averaging similar and the orientation of the major axis of the ellipse. In [27], a

380
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

model-based feature analysis method was presented for the − 1, j= 0, . . . , N – 1. Where M, N denote the number of
automatic extraction and description of human gait for rows and columns of the silhouette, respectively.
recognition. The method generated a gait signature using a
Fourier series expansion of a signal corresponding to the hip Let s [i , j ] = ⎧⎨ 1 if (i, j) belongs to the foreground
rotation [Figure 3(c)]. In [28], a more detailed model was ⎩0 otherwise
proposed using ellipses for the torso and the head, line Using the above term, the horizontal and vertical
segments for the legs, and a rectangle for each foot [Figure projection of silhouettes [17] are expressed as
3(d)].
p h [i ] = ∑ s [ i , j ],i = 0 ,1,..... M − 1
N −1
(1)
j=0

∑
M −1
pv [ j] = s [ i , j ], j = 0 ,1,...... N − 1 (2)
i=0

The efficiency of this feature is based on the fact that it is

sensitive to silhouette deformations since all pixel
movements are reflected in the horizontal or vertical
Fig. 3. Graphical Representation of parameters used in model-based projection [Figure 4(b)]. Although this feature is similar to
approaches. (a) Distance used as static parameters in [26], (b) Ellipse fitting the width of silhouette (note the similarity between the
in silhouette regions[22], (c) Hip rotation model [27] and (d) Model using a
combination of shapes[28].
width vector and the horizontal projection vector), it is more
robust to spurious pixels. An angular transform of the silhouette
VII. MODEL FREE APPROACHES was proposed in [20].The angular transform divides the
silhouettes into angular sectors and computes the average
Model free solutions operate directly on the gait distance between foreground pixels and the center (ic, jc) of
sequences without assuming any specific model for the the silhouette [Figure 4(c)].
walking human. A very interesting class of holistic
1
techniques merely employs binary maps (silhouettes) of A (θ ) =
Nθ
∑ ( i , j )∈ Fθ
s[ i , j ] ( i − ic ) 2 + ( j − j c ) 2 (3)
walking humans. Such techniques are particularly suited for
most practical applications since color or texture where, θ is an angle, Fθ is the set of the pixels in the circular
information may not be available or extractable. The contour sector [θ − ( Δ θ / 2 ), θ + ( Δ θ / 2 ) ] and is the
of the silhouette is probably the most reasonable feature in
this class. It can be used directly [19], or it can be cardinality of Fθ . The transform coefficients were shown to
transformed to extract Fourier descriptors [29].The width of be a linear function of the silhouette contour.
silhouette was proposed in [30] as a suitable feature for gait
feature extraction. The width w[i] of silhouette is the
horizontal distance between the leftmost and rightmost
foreground pixels in each row i of the silhouette [Figure
4(a)]. Although the calculation of width signals imposes
minimal processing load on a gait system, algorithms that
use this feature are vulnerable to spurious pixels that often
render the identification of the leftmost and rightmost pixels
inaccurate. For this reason, the authors in [30] propose a Fig. 4. Features extracted from binary silhouettes for gait recognition (a)
Width of silhouette, (b) Vertical and Horizontal Projections and (c) Angular
postprocessing technique to smooth and denoise the feature
representation.
vectors prior to their deployment in gait recognition.
Henceforth, we assume that each gait sequence is composed
of several binary silhouettes, denoted as s[i, j], i = 0, . . . , M
TABLE I: ADVANTAGES & DISADVANTAGES OF THE HOLISTIC AND MODEL BASED APPROACHES
HOLISTIC APPROACH
FEATURE ADVANTAGES DISADVANTAGES
CONTOUR Sensitive to Structural Differences High Complexity, Low Robustness
WIDTH Sensitive to Structural Differences, Low Complexity Low Robustness
PROJECTIONS Robustness, Low Complexity Coarse Structural Representation
ANGULAR Robustness Coarse Structural Representation
RELATIVE PHASES Compact Representation, Scale Invariance Complicated Determination of phases
SILHOUETTE Lossless Representation Leads to high-complexity systems
MODEL BASED APPROACH
FEATURE ADVANTAGES DISADVANTAGES
STATIC PARAMETERS View Invariant, Compact Representation Difficult Capturing
ELLIPSE PARAMETERS Compact Representation Low Robustness
HIP ANGLE Compact Representation Low Robustness
COMPBINATION OF Compact Representation Low Robustness
SHAPE PARAMETERS

381
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

The silhouette itself was used in several algorithms as a On the temporal axis, it appears that shape information
feature. Prior to their deployment, the silhouettes in a gait can be captured using four or five characteristic frames [7],
sequence should be appropriately scaled and aligned. In [24] or feature vectors. Since several of the elements in the
most cases, it appears that the silhouette is at least as feature vectors, extracted using the techniques in the
efficient as the low-dimensional features that can be previous sections, usually contain information that does not
extracted from a silhouette. contribute to the purpose of recognition, methodologies such
as principal component analysis (PCA) [30], [16], [19] or
linear discriminant analysis (LDA) [16] are used to retain
VIII. FREQUENCY TRANSFORMATION OF FEATURE TIME only the important elements of the original feature vector.
SERIES Analysis of variance (ANOVA) can also be used for the
Since walking is a periodic activity, the Fourier analysis identification of the significant components in a gait feature
of the time-domain gait signals is a very appealing approach vector. Several works achieve good performance using
as most discriminative information is expected to be holistic features of dimension as low as 100. On the other
compacted in a few Fourier coefficients, providing a very hand, feature vectors consisting of model parameters would
efficient gait representation. Therefore, taking the Fourier carry more information than feature vectors extracted using
transform of the feature vector series f(t). a holistic method.
2π
1 −j Kt
∑
T −1
F (K ) = t =0
f (t ) e T (4)
X. PATTERN MATCHING AND CLASSIFICATION
T
Once gait information is extracted from gait sequences
where T is the walking period, yields a new representation
and the associated feature vectors are formed, the actual
that is related to the frequency content of the originally
recognition/classification task must be performed. Two main
extracted features. Specifically, since the transform is
approaches can be taken, namely, a template-based approach
calculated in increments of the angular frequency 2π/T,
or a stochastic approach. In both cases, an appropriate
signals extracted from gait sequences with different walking
distance metric between feature vectors must be initially
periods are directly comparable. In practice, however, not all
defined. The classical Euclidean distance is the measure that
frequency components are useful for recognition. This is
is used in most gait recognition applications. Other measures
why there are methods that use only the magnitude and
are the inner product distance [24] and the number of “ones”
phase of the Fourier transform at the fundamental walking
in the binary difference between frames [15]. A variety of
frequency [22], [31]. In [31], it was stated that all motions in
other distance measures may also be used [32]. However, in
a gait cycle share the same fundamental frequency, and a
this work, we use the classical Euclidean distance in the
system was proposed which uses optical flow for measuring
implementations of the presented gait methodologies.
shape oscillations. A significant conclusion reached in [22]
was that frequency signatures yielded superior performance TABLE II: IMPACT ON AVERAGE RESULTS DUE TO DIFFERENCE IN
in cases where the compared gait sequences were captured CAPTURING CONDITIONS
on different days (and, therefore, the structural information PROBE DIFFERENCE RANK-1 RANK-5
alone was not reliable). This provides an additional motive A View 90 98
for investigating frequency- domain features. In the B Shoe 80 87
experimental assessment section, we will evaluate the C Shoe, View 68 83
performance of a simple scheme based on the direct
D Surface 25 53
transformation of features. In this system, the entire feature
E Surface, Shoe 21 57
time series is expressed as a single complex feature vector
F Surface, View 18 51
through application of. In Figure 6, we display such a
representation using silhouettes. G Surface, Shoe, View 18 46

XI. TEMPLATE MATCHING

The main concern in calculating distances between
different gait representations (templates) is whether we
compare corresponding quantities in the two representations.
In case of frequency templates (e.g., harmonic components
Fig. 5. Frequency signatures: a complex silhouette template computed as computed using Fourier analysis), the calculation of the
the Fourier transform of the gait sequence at the fundamental frequency. (a) distance between two templates is straightforward since the
Real part and (b) imaginary part. correspondence between frequency components in different
templates is obvious. In the case of spatial templates, the
gait representation is a sequence of features that must be
IX. DIMENSIONALITY REDUCTION
compared with another sequence of features. When the
A natural question arises in the context of gait analysis: fundamental walking periods T1 and T2 of the two sequences
How much information do we need to extract from a gait are not equal, their cumulative distance over a gait cycle is
sequence in order to capture most discriminative defined as
information?

382
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

1 The HMM-based methodology is, in many aspects,

∑t =1 u (t ) D( f1 (w1 (t )), f 2 (w2 (t ))) ,
T
D12 = preferable to other techniques since it explicitly takes into
U
consideration not only the similarity between shapes in the
where the pairs (w1(t ),w2(t )) define a warping function, u(t) test and reference sequences, but also the probabilities with
∑
T
is a weighting function, U = u (t ) , and D(·) denotes which shapes appear and succeed each other in a walking
t =1
cycle of a specific subject.
the distance between the feature vectors at time t. Based on
the characteristics of the warping function, we can
distinguish three approaches for the calculation of distances XIII. EXPERIMENTAL ASSESSMENT
between feature sequences. The direct matching approach
To evaluate the efficiency of the main gait analysis and
can be regarded as a brute-force attempt to match a pattern
recognition approaches that were presented previously, we
consisting of feature vectors (derived from frames in a gait
considered several features, as well as both the template
cycle) by sliding it over a sequence of feature vectors of the
reference sequence to find the position that yields the matching and statistical approaches for the recognition stage.
minimum distance. This is the approach taken in the Although there are several gait databases for the evaluation
baseline method created at USF [15]. The use of time of the main approaches, as summarized in Table III, we used
normalization [33] is a more reasonable approach since the USF database, which is used by most researchers in the
reference and test sequences corresponding to the same gait community for reporting results. Prior to testing, we
subject may not necessarily have the same gait period. aligned the silhouettes to the center of the frames to give a
Consequently, if recognition is to be performed by template fair comparison with features that are not translation-
matching, some kind of compensation would have to be invariant. We tested several features and recognition
applied during the calculation of the distance. To this end, methods. We considered only holistic features here since
dynamic time warping (DTW) [33] can be used to calculate they generally outperform the model-based features and they
the distance between a test sequence and a reference are more interesting from a signal processing perspective.
sequence. Using DTW [30], [25], all distances between test For the evaluation of the efficiency of features, we formed
and reference frames are computed and the total distance is
feature vectors of appropriate size. The size of the width
defined as the accumulated distance along the minimum-
vector [30] was equal to the vertical dimension of the
distance path (termed the optimal warping path). Another
silhouettes. The width vector was filtered with a three-tap
option is to use linear time normalization. Having computed
the distances between a test subject and all subjects in a low-pass filter since this approach was reported to yield
reference database, the recognition decision is taken as better results. The projections vector [17] was generated as a
concatenation of the horizontal and vertical projections and,
identity(i ) = arg min j Dij therefore, its size was set equal to the sum of the horizontal
and vertical dimensions of the silhouettes. For the angular
where Dij denotes the cumulative distance between the ith
feature, we calculated the transform coefficients in circular
test subject and the jth reference subject. This means that the
sectors of 5◦. This yielded 72-dimensional feature vectors.
identity of the test subject is assumed to be the identity of
In this section, we report results in terms of cumulative
the reference subject with which the test subject has the
minimum distance. match scores. To calculate these scores, we conduct multiple
tests using multiple test sequences. Each test sequence is
compared to the sequences in the reference database (for
XII. STATISTICAL APPROACH: HMMS each test sequence there is only one correct match in the
Stochastic approaches such as HMMs can also be used for reference database), and the sequences in the reference
gait recognition [24]. In practical HMM-based gait database are ranked according to their similarity with the
recognition, each walking subject is assumed to traverse a test sequence. As proposed in [36], rank-n performance is
number of stances. In other words, each frame in a gait calculated by measuring the percentage of tests in which the
sequence is considered to be emitted from one of a limited correct subject appears in the top n matches. The results,
number of stances. The a priori probabilities, as well as the
rank-1 and rank-5 scores averaged over all test sets in the
transition probabilities, are used to define models λ for each
gait challenge database, are tabulated in Table IV. It is seen
subject in a reference database. For a test sequence of
feature vectors, the probability that it was generated by one that features that do not depend on the detection of boundary
of the models associated with the database sequences can be pixels offer the best performance. Despite the fact that the
calculated by database on which the features were tested was quite noisy,
experiments on less noisy conditions demonstrate that these
P ( f i / λ j ), j = 1,2...N features would still be superior, occasionally with a
narrower margin.
where N is the number of subjects in the reference database.
In any case, the noisy conditions should be considered as
The subject corresponding to the model yielding the higher
probability is considered to be identical to the test subject, the general rule in gait recognition since only in laboratory
i.e. environments is it possible to achieve perfectly clean
silhouettes. All features were combined and tested with
identity(i ) = arg max j P ( f i / λ j ), j = 1,....N
several recognition methodologies.

383
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

TABLE III: LIST OF DATABASES USED FOR GAIT RECOGNITION.

DATA URL SUBJECTS PARAMETERS
USF/NIST www.gaitchallenge.org 71 Viewpoint, Surface, Shoe
CMU www.hid.ri.cmu.edu 25 Viewpoint, Walking Speed, Carried Object
SOTON www.gait.ecs.soton.ac.uk 118 Viewpoint, Treadmill, Indoors/Outdoors
SHAPE OF MOTION Pages.cpsc.ucalgary.ca/~boyd/gait/gait.html 6 -
GATECH www.cc.gatech.edu/cpl/projects/hid/ 20 Viewpoint, Indoors/Outdoors
CASIA www.sinobiometrics.com 20 Viewpoint
MIT www.ai.mit.edu/projects/gait/ 25 Time

TABLE IV: RESULTS OBTAINED FOR SEVERAL COMBINATIONS OF FEATURES & RECOGNITION METHODS OVER ALL SETS OF GAIT DATABASE CLASSIFIED
INTO RANK-1 (R1) AND RANK-5(R5).
FEATURE WIDTH PROJECTION ANGULAR SILHOUETTE AVERAGE
R1 R5 R1 R5 R1 R5 R1 R5 R1 R5
Frequency Domain Distance 21 42 26 45 20 41 46 66 28 49
Dynamic Time Wrapping 26 49 33 56 36 59 47 70 36 59
Linear Time Normalization 28 49 35 55 36 60 46 74 36 59
Hidden Markov Models 34 51 36 49 36 62 45 70 38 58
Structural Matching 33 50 35 47 36 60 43 62 37 55
Average 28 48 33 50 33 56 45 68 - -

Although the frequency signatures constitute features that, in the case of gait, recognition using these methods
derived from time-domain features, here we treat the seems to be based predominantly on structural similarities
frequency-domain approach as a recognition method that between compared sequences and/or that the gait dynamics
computes the Fourier transform of the feature time series at can be captured equally well by the linear and the nonlinear
the walking frequency and compares the resultant normalization processes.
components using a Euclidean distance. For DTW and linear The results derived using the structural matching
time normalization, the distance between test and reference approach discloses the importance of shape in gait
gait sequences was computed by taking the median of the recognition. We see that, although gait dynamics are ignored,
distances from all combinations between cycles in the with this approach the performance of the system is
reference and test gait sequences. For testing the generally good in comparison to the rest of the approaches.
performance of HMMs for gait recognition, we It also reinforces our belief that current approaches for gait
implemented the algorithm in [35]. In the structural recognition primarily depend on structure rather than on gait
matching approach, we computed for each subject the dynamics. The performance of the system deploying HMMs
minimum cumulative distance of gait frames to the is better than that achieved using structural matching.
exemplars determined using the HMM model. This However, the performance gain is not very impressive, and
experiment was intended to show how much of the this makes us believe that there might be other more
performance of the HMM approach is due to the appropriate ways to exploit gait dynamics.
computation of structural similarities and how much is due
to the exploitation of gait dynamics. Complete cumulative
match score curves are shown in Figure 6 for recognition XIV. CONCLUSION
based on frequency signatures, DTW, linear time This article was intended to provide an overview of the
normalization, HMMs, and structural matching. As seen, the basic research directions in the field of gait analysis and
frequency signature approach is quite efficient despite the recognition. The recent developments in gait research
fact that it is the least complicated of all approaches in our
indicate that gait technologies still need to mature and that
comparison. Since the determination of similarity between
limited practical applications should be expected in the
frequency signatures is direct, i.e., there is no need to find
immediate future. At present, there is a potential for initial
the correspondences in two compared frequency
representations, the savings in computational complexity is deployment of gait for recognition in conjunction with other
considerable and the approach appears to be rather biometrics. However, future advances in gait analysis and
appealing. recognition an open, challenging research area—are
In general, the DTW and the linear time normalization expected to result in wide deployment of gait technologies
approaches perform roughly the same. As mentioned in the not only in surveillance, but in many other applications as
previous section, this is a rather unexpected result since, in well. We hope that this article will expose the gait analysis
the context of speech recognition, it was reported [34] that and recognition problem to the signal processing community
DTW performed clearly better than linear time and that it will stimulate the involvement of more
normalization. A possible explanation might lie in the fact researchers in gait research in the future.

384
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

Fig. 6. Cumulative match scores for different recognition approaches using the silhouette feature. (a) Frequency signature, (b) dynami time warping, (c)
linear time normalization, (d) structural matching, and (e) hidden Markov models.

[7] R.T. Collins, R. Gross, and J. Shi, “Silhouette-based human

identification from body shape and gait,” in IEEE Conference
REFERENCES Automatic Face and Gesture Recognition, Washington, DC, May
2002, pp.351–356.
[1] L. Sloman, M. Berridge, S. Homatidis, D. Hunter, and T. Duck, “Gait [8] G. Johansson, “Visual perception of biological motion and a model
patterns of depressed patients and normal subjects,” American Journal for its analysis,” Percept. Psycophysics, vol. 14, no. 2, pp. 201–211,
of Psychiatry, vol. 139, no. 1, pp. 94–97, 1982. 1973.
[2] M.D. Addlesee, A. Jones, F. Livesey, and F. Samaria, “ORL active [9] J.E. Cutting and L.T. Kozlowski, “Recognizing friends by their walk:
floor,” IEEE Personal Communication, vol.4,no. 5, pp. 35–41, Oct. Gait perception without familiarity cues,” Bulletin Psychonomic Soc.,
1997. vol. 9, no. 5, pp. 353–356, 1977.
[3] J. Kittler, M. Hatef, R.P.W. Duin, and J. Matas, “On combining [10] S.A. Niyogi and E.H. Adelson, “Analyzing and recognizing walking
classifiers,” IEEE Transaction Pattern Analysis Machine Intelligence, figures in XYT,” in Proc. Computer Vision and Pattern Recognition,
vol. 20, no. 3, pp. 226–239, Mar. 1998. Seattle, WA, June 1994, pp. 469–474.
[4] G. Shakhnarovich, L. Lee, and T. Darrell, “Integrated face and gait [11] H. Murase and R. Sakai, “Moving object recognition in eigenspace
recognition from multiple views,” in Proceeding IEEE Conference representation: Gait analysis and lip reading,” Pattern Recognition
Computer Vision and Pattern Recognition, Kauai, HI, Dec. 2001, vol. Letter, vol. 17, no. 2, pp. 155–162, 1996.
I, pp. 439–476. [12] D. Cunado, M.S. Nixon, and J.N. Carter, “Using gait as a biometric,
[5] A. Kale, A.K. RoyChowdhury, and R. Chellappa, “Fusion of gait and via phaseweighted magnitude spectra,” in Proc. International
face for human identification,” in Proceeding IEEE International Conference Audio- and Video-Based Biometric Person
Conference Acoustics, Speech, and Signal Processing, Montreal, Authentication, Crans-Montana, Switzerland, Mar. 1997, vol. LNCS
Canada, May 2004, vol. 5, pp. 901–904. 1206, pp. 95–102.
[6] E. Ayyappa, “Normal human locomotion, part 1: Basic concepts and [13] J. Little and J. Boyd, “Recognizing people by their gait: The shape of
terminology,” Journal of Prosthetics Orthotics, vol.9,no.1, pp.10–17, motion,” Videre, Int. Journal of Computer Vision, vol. 14, no. 6, pp.
1997. 83–105, 1998.

385
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

[14] P.S. Huang, C.J. Harris, and M.S. Nixon, “Visual surveillance and [37] M. Murray. Gait as a total pattern of movement. American Journal of
tracking of humans by face and gait recognition,” in Proc. 7th IFAC Physical Medicine, vol. 46(1):pp. 290–333. ISSN 0002-9491.
Symposium Artificial Intelligence in Real-Time Control, Grand February 1967.
Ganyon National Park, AZ, Oct. 1998, pp. 43–44. [38] J. E. Cutting and L. T. Kozlowski. Recognizing Friends by their Walk:
[15] P.J. Phillips, S. Sarkar, I. Robledo, P. Grother, and K.W. Bowyer, Gait Perception without Familiarity Cues. Psychonomic Society, vol.
“The gait identification challenge problem: Data sets and baseline 9(5):pp. 353–6. 1977.
algorithm,” in Proc. International Conference Pattern Recognition, [39] M. Addlesee, A. Jones, F. Livesey, and F. Samaria. The ORL Active
Quebec City, Canada, Aug. 2002, vol. 1, pp. 385–388. Floor. IEEE Personal Communications, vol. 4(5):pp. 35–41. 1997.
[16] C. BenAbdelkader, R. Cutler, and L. Davis, “Motion-based [40] J. Little and J. Boyd. Recognizing People by their Gait: the Shape of
recognition of people in eigengait space,” in Proc. IEEE Int. Conf. Motion. Videre, vol. 1(2). 1998.
Automatic Face and Gesture Recognition, Washington, DC, May [41] P. Huang, C. Harris, and M. Nixon. Comparing Different Template
2002, pp. 254–259. Features for Recognizing People by their Gait. In Lewis and Nixon
[17] Y. Liu, R. Collins, and Y. Tsin, “Gait sequence analysis using Frieze [Lewis98], pp. 639–48. September 1998.
patterns,” in Proc. European Conference Computer Vision, [42] P. Huang, C. Harris, and M. Nixon. Recognizing Humans by Gait
Copenhagen, May 2002, vol. LNCS 2351, pp. 657–671. Using a Statistical Approach for Temporal Templates. In SMC’98
[18] J.P. Foster, M.S. Nixon, and A. Prugel-Bennett, “Automatic gait Conference Proceedings. 1998 IEEE International Conference on
recognition using area-based metrics,” Pattern Recognition Letter, Systems, Man, and Cybernetics, vol. 5. IEEE. October 1998.
vol. 24, no. 14, pp. 2489–2497, 2003. [43] P. S. Huang, C. J. Harris, and M. S. Nixon. Canonical Space
[19] L. Wang, T. Tan, H. Ning, and W. Hu, “Silhouette analysis-based gait Representation for Recognizing Humans by Gait and Face. In IEEE
recognition for human identification,” IEEE Trans. Pattern Analysis. Southwest Symposium on Image Analysis and Interpretation. IEEE.
Machine Intelligence, vol. 25, no. 12, pp. 1505–1518, Dec. 2003. April 1998.
[20] N.V. Boulgouris, K.N. Plataniotis, and D. Hatzinakos, “An angular [44] P. S. Huang, C. J. Harris, and M. S. Nixon. A Statistical Approach for
transform of gait sequences for gait assisted recognition,” in Proc. Recognizing Humans by Gait Using Spatial-Temporal Templates. In
IEEE Int. Conf. Image Processing, Singapore, Oct. 2004, pp. 857–860. International Conference on Image Processing, vol. 3. October 1998.
[21] L. Lee, G. Dalley, and K. Tieu, “Learning pedestrian models for [45] P. Huang, C. Harris, and M. Nixon. Human Gait Recognition in
silhouette refinement,” in Proc. Int. Conf. Computer Vision, Nice, Canonical Space using Temporal Templates. Vision, Image and
France, Oct. 2003, pp. 663–670. Signal Processing, vol. 146(2):pp. 93– 100. April 1999.
[22] L. Lee and W.E.L. Grimson, “Gait analysis for recognition and [46] J. M. Nash, J. N. Carter, and M. S. Nixon. Extraction of Moving
classification,” in Proc. IEEE Int. Conf. Automatic Face and Gesture Articulated-Objects by Evidence Gathering. In Lewis and Nixon
Recognition, Washington, DC, May 2002, pp. 148–155. [Lewis98], pp. 609–18. September 1998.
[23] D. Tolliver and R.T. Collins,“Gait shape estimation for [47] D. Meyer, J. P¨osl, and H. Niemann. Gait classification with HMMs
identification,” in Proc.4th Int. Conf. Audio and Video-Based for trajectories of body parts extracted by mixture densities. In Lewis
Biometric Person Authentication, Guilford, UK, June 2003, pp.734– and Nixon [Lewis98], pp. 459–68. September 1998.
742. [48] R. Orr and G. Abowd. The smart floor: A mechanism for natural user
[24] A. Sundaresan, A.K. Roy Chowdhury, and R. Chellappa, “A hidden identification and tracking. In Conference on Human Factors in
Markov model based framework for recognition of humans from gait COmputing Systems. 2000.
sequences,” in Proc. Int. Conf. Image Processing 2003, Barcelona, [49] A. F. Bobick and A. Johnson. Gait recognition using static activity-
Spain, Sept. 2003, vol. 2, pp. 14–17. specific parameters. In IEEE Conference on Computer Vision and
[25] N.V. Boulgouris, K.N. Plataniotis, and D. Hatzinakos, “Gait Pattern Recognition. Dec 2001.
recognition using dynamic time warping,” in Proc. IEEE Int. Symp. [50] A. Y. Johnson and A. F. Bobick. A Multi-view Method for Gait
Multimedia Signal Processing, Siena, Italy, Sept. 2004, pp. 263–266. Recognition Using Static Body Parameters. In International
[26] A. Johnson and A. Bobick, “A multi-view method for gait recognition Conference on Audio- and Video Based Biometric Person
using static body parameters,” in 3rd Int. Conf. Audio- and Video- Authentication. June 2001.
Based Biometric Person Authentication, Halmstad, Sweden, June [51] R. Gross and J. Shi. The CMU Motion of Body (MoBo)Database.
2001, pp. 301–311. Tech. Rep. CMU-RI-TR-01-18, Robotics Institute,Carnegie Mellon
[27] D. Cunado, M.S. Nixon, and J.N. Carter, “Automatic extraction and University. June 2001.
description of human gait models for recognition purposes,” Comput. [52] J. P. Phillips, S. Sarkar, I. Robledo, P. Grother, and K. W. Bowyer.
Vis. Image Understand., vol. 90, no. 1, pp. 1–14, 2003. Baseline Results for the Challenge Problem of Human ID Using Gait
[28] D.K. Wagg and M.S. Nixon, “On automated model-based extraction Analysis. To appear at International Conference on Automatic Face
and analysis of gait,” in Proc. IEEE Int. Conf. Automatic Face and and Gesture Recognition. May 2002.
Gesture Recognition, Seoul, Korea, May 2004, pp. 11–16.
[29] S.D. Mowbray and M.S. Nixon, “Automatic gait recognition via
Fourier descriptors of deformable objects,” in Proc. Int. Conf. on Praveen Kumar Gupta born on January 20-1979 at
Audio- and Video-Based Biometric Person Authentication, Guilford, Kanpur,U.P. in india. And received the Master of
UK, June 2003, pp. 566–573. Computer Application Degree (MCA) Agra
[30] A. Kale, N. Cuntoor, B. Yegnanarayana, A.N. Rajagopalan, and R. University, Agra (Uttar Pradesh, India) in 2002. He
Chellappa, “Gait analysis for human identification,” in 4th Int. Conf. received Master of Business Administrator Degree
Audio- and Video-based Person Authentication, Guilford, UK, June (MBA) in System and Marketing from SMU
2003, pp. 706–714. University, Manipal (Sikkim, India) in 2007.
[31] J.E. Boyd and J.J. Little, “Phase models in gait analysis,” presented at He is currently working as assistant professor at
the Workshop on Models Versus Exemplars at Computer Vision and Praveen Singh Institute of Technology (PSIT), Kanpur ( U.P., India). He
Pattern Recognition, Kauai, HI, Dec. 2001. has 9 years teaching experience. He guided 10 under graduate projects and
[32] K.N. Plataniotis, D. Androutsos, and A.N. Venetsanopoulos, 10 post graduate projects.
“Adaptive fuzzy systems for multichannel signal processing,” Proc.
IEEE, vol. 87, no. 9, pp. 1601–1622, Sept. 1999. Rajendra Singh born on July 25-1981 at Kanpur,U.P.
[33] H. Sakoe and S. Chiba, “Dynamic programming algorithm in india. And received the Bachelor’s Degree (B.
optimization for spoken word recognition,” IEEE Trans. Acoustics,
Tech.) in Computer Science and Engineering from
Speech, Signal Processing, vol. 26, no. 1, pp. 43–49, Feb. 1978.
University Institute of Engineering and Technology
[34] H. Sakoe and S. Chiba, “Dynamic programming algorithm
(UIET), Chhatrapati Shahu Ji Maharaj University,
optimization for spoken word recognition,” IEEE Trans. Acoustics,
Speech, Signal Processing, vol. 26, no. 1, pp. 43–49, Feb. 1978. Kanpur (Uttar Pradesh, India) in 2005. He received
[35] A. Kale, A. Sundaresan, A.N. Rajagopalan, N. Cuntoor, A.K. Roy- Master’s Degree (M. Tech) in Computer Engineering
Chowdhury, V. Krueger, and R. Chellappa, “Identification of humans from Institute of Technology and Management (ITM),
using gait,” IEEE Trans. Image Processing, vol. 13, no. 9, pp. 1163– Maharshi Dayanand University, Rohtak (Haryana,
6173, Sept. 2004. India) in 2009.
[36] P.J. Phillips, H. Moon, S. Rizvi, and P. Raus, “The feret evaluation He is currently working as assistant professor at Praveen Singh Institute
methodology for face recognition algorithms,” IEEE Trans. Pattern of Technology (PSIT), Kanpur (U.P., India). He has 4.5 years teaching
Anal. Machine Intell. vol. 22, no. 10, pp. 1090–1104, 2000. experience. He guided 5 under graduate projects and 2 post graduate
projects.

386
 International Journal of Machine Learning and Computing, Vol. 1, No. 4, October 2011

Rohit Katiyar born on April-13-1983 at Varanasi, Ritesh Rastogi born on January-21-1974 at Dehra
U.P. in India. and received the B. Tech. (Hons.) Dun Uttaranchal in India. And received the MCA
degree in Information Technology from B.S.A. degree from MDU Rohtak University He is currently
College of Engineering & Technology affiliated with working toward the Ph.D. degree at Manav Bharti
U. P. Technical University, Lucknow in 2005. He is Uiversity,Solan. His current research interests
currently working toward the Ph.D. degree at include Image Processing and Fusion of different
Harcourt Butler Technological Institute, Kanpur. His biometrics modalities etc
current research interests include Gait Biometrics He is currently working as Astt. Prof at Noida
System, Image Processing and Fusion of different Institute of Engineering and Technology Greater
biometrics modalities etc. Noida. He has 11 years teaching experience.
He is currently working as teacher fellow at Harcourt Butler Mr.Rastogi is also a member of ISTE society. Currently Mr. Rastogi has
Technological Institute, Kanpur. He has 5.5 years teaching experience and 1 international conference papers, 1 international and 6 national journal. He
4 years experience in research. He is also a counselor in IGNOU study has guided 20 undergraduate projects and 25 postgraduate projects.
centre from last 4 years.
Mr. Katiyar is also a member of IACSIT society and also one of the peer
reviewer member in IACSIT. Currently Mr. Katiyar has 6 international
conference papers, 1 international and 1 national journal. He has guided 10
undergraduate projects and 15 postgraduate projects.

387