28-03-2012, 03:58 PM
AUTOMATIC GAIT RECOGNITION
AUTOMATIC GAIT RECOGNITION.pdf (Size: 225.18 KB / Downloads: 61)
Abstract
Gait is an emergent biometric aimed essentially to
recognise people by the way they walk. Gait’s advantages are that
it requires no contact, like automatic face recognition, and that it
is less likely to be obscured than other biometrics. Gait has allied
subjects including medical studies, psychology, human body
modelling and motion tracking. These lend support to the view
that gait has clear potential as a biometric. Essentially, we use
computer vision techniques to derive a gait signature from a
sequence of images. The majority of current approaches analyse
an image sequence to derive motion characteristics that are then
used for recognition; only one approach is feature based. Early
results by these studies confirm that there is a rich potential in
gait for recognition. Only continued development will confirm
whether its performance can equal that of other biometrics and
whether its application advantages will indeed make it a
pragmatist’s choice.
Key words: Gait, walking, biometrics.
2 Nixon et al.
without being obvious in order not to provoke attention. On escape, again the robber
will either exit at speed, or in (apparent) leisure. The motion in both cases is natural,
for the subject will either not want to attract attention or to move quickly.
Clearly, there are limits to the use of gait as a biometric, a detailed study of the
limitations awaits development of technique. However, it is not unlikely that
footwear can affect gait, as can clothing. Equally, physical condition can affect gait
such as pregnancy, affliction of the legs or feet, or even drunkenness. These factors
are not new to biometrics: a face can be made up or have spectacles, ears can be
obscured by hair, hands can even be cut off, as acknowledged in other chapters. As
usual, a major question concerns whether these are part of human perception whereas
a biometric system can perceive the underlying characteristics of the biometric - in the
case of gait, the individual’s musculature which essentially limits the variation of
motion. As such, these factors await investigation.
The view that gait can be used to recognize individual is not new: Shakespeare
used a rich lexicon of adjectives to describe gait, including princely, lion’s, heavy,
humble, weary, forced, gentle, swimming, and majestic. Further, in The Tempest [Act
4 Scene 1], Ceres observes
“High’st Queen of state, Great Juno comes; I know her by her gait”.
Even more, in Troilius and Cressida [Act 4 Scene 5], Ulysses states
“Tis he, I ken the manner of his gait; He rises on the toe: that spirit of his in
aspiration lifts him from the earth”.
The former is one of Shakespeare’s many observations on recognizing people by their
gait; the latter includes a concise description of Diomedes’ demeanour.
Accordingly, there appears much potential for using gait as a biometric. There
have been allied studies, particularly those in medical studies for therapy, but there
have also been psychological studies, and approaches aimed to model and track
human targets through an image sequence, though not usually for recognition, as
discussed in Section 2. Current approaches to automatic gait recognition are surveyed
in Section 3, together with a more detailed examination of two extant approaches to
automatic recognition. Possibilities for further work are discussed in Section 4 prior to
the conclusions concerning the potential for gait as a biometric.
2. Allied Research
Medical Studies
The aim of medical research has been to classify the components of gait for the
treatment of pathologically abnormal patients. Murray et al. [34] produced standard
movement patterns for pathologically normal people which were used to compare the
gait patterns for pathologically abnormal patients [35]. The data collection system
used required markers to be attached to the subject. This is typical of most of the data
collection systems used in the medical field, and although practical in that domain,
they are not suitable for identification purposes. Gait was considered by Murray as “a
total walking cycle” - the action of walking can be thought of as a periodic signal. The
following terms are used to describe the gait cycle, as given in [34], and are used
throughout the report. Fig. 11.1 illustrates the terms described. A gait cycle is the time
Automatic Gait Recognition 3
interval between successive instances of initial foot-to-floor contact ‘heel strike’ for
the same foot. Each leg has two distinct periods; a stance phase, when the foot is in
contact with the floor, and a swing phase, when the foot is off the floor moving
forward to the next step. The cycle begins with the heel strike of one foot which
marks the start of the stance phase. The ankle flexes to bring the foot flat on the floor
and the body weight is transferred onto it. The other leg swings through in front as the
heel lifts of the ground. As the body weight moves onto the other foot, the supporting
knee flexes. The remainder of the foot, which is now behind lifts off the ground
ending the stance phase.
Figure 11.1 Relationship between temporal components of the walking cycle and the
step and stride lengths during the cycle.
The start of the swing phase is when the toes of the foot leave the ground. The
weight is transferred onto the other leg and the leg swings forward to meet the ground
in front of the other foot. The gait cycle ends with the heel strike of the foot. Stride
length is the linear distance in the plane of progression between successive points of
contact of the same foot. Step length is the distance between successive contact points
of opposite feet. A step is the motion between successive heel strikes of opposite feet;
a complete gait cycle is comprised of two steps.
Murray et al.’s work [34,35] suggests that if all gait movements were considered,
gait is unique. In all there appear to be twenty distinct gait components, some of
which can only be measured from an overhead view of the subject. Murray found “the
pelvic and thorax rotations to be highly variable from one subject to another” [35].
These patterns would be difficult to measure even from an overhead view of the
subject, which would not be suited to application in many practical situations. Murray
also suggested that these rotation patterns were not found to be consistent for a given
individual in repeated trials. In [34,35] ankle rotation, pelvic tipping and spatial
displacements were shown to possess individual consistency in repeated trials.
Unfortunately, these components would be difficult to extract from real images.
The normal hip rotation pattern of the angle of the thigh (the angle between the
thigh and horizontal) is characterized by one period of extension and one period of
4 Nixon et al.
flexion in every gait cycle, as described by Murray. Fig. 11.2 gives the average
rotation pattern: the upper and lower lines indicate the standard deviation from the
mean. In the first half of the gait cycle, the hip is in continuous extension as the trunk
moves forward over the supporting limb. In the second phase of the cycle, once the
weight has been passed onto the other limb, the hip begins to flex in preparation for
the swing phase. This flexing action accelerates the hip, directing the swinging limb
forward for the next step. Later, we will see how these angles have featured in a
model-based recognition system.
There is an extensive literature on studies of gait for medical use, none of which is
concerned primarily with biometrics. Intuitively, measurements by gait researchers
could prove to be of benefit in biometrics, though there is natural concern that the
markers used do not realistically capture individual characteristics. Using gait as a
biometric concerns its derivation by computer vision, for this is the only way it can
satisfy its purpose. Some insight into gait as a biometric can however be drawn from
psychology.
Figure 11.2 Mean hip rotation pattern [35].
Psychology of Gait
In the earliest studies of gait perception [21] participants were presented with images
produced from points of light attached to body joints. When the points were viewed in
static images, they were not perceived to be in human form, rather that they formed a
picture - of a Christmas tree even. When the points were animated, they were
immediately perceived as representing a human in motion. Later work showed how by
point light displays a human could be rapidly extracted and that different types of
motion could be discriminated, including jumping and dancing [15]. More recently
Bingham [5] has shown that point light displays are sufficient for the discrimination
of different types of object motion and that discrete movements of parts of the body
can be perceived. As such, human vision appears adept at perceiving human motion,
even when viewing a display of light points. Indeed, the redundancy involved in the
Automatic Gait Recognition 5
light point display might provide an advantage for motion perception [39] and could
perhaps offer improved performance over video images.
Naturally, studies in perception have also addressed gender as well as pure motion,
again using point light displays. One early study [25] showed how gender could be
perceived, and how accuracy was improved by inclusion of height information [44].
The ability to perceive gender has been attributed to anatomical differences which
result in greater shoulder swing for men, and more hip swing for women [30]. Indeed,
a torso index (the hiphoulder ratio) has been shown to discriminate gender [14], and
the identification of gender by motion of the center of moment was also suggested.
Gender identification would appear to be less demanding than person
identification. However, it has been shown that subjects could recognize themselves
and their friends [12], and this has been explained by considering gait as a
synchronous, symmetric pattern of movement from which identity can be perceived
[13]. Like Shakespeare’s observations, these studies encourage the view that gait can
indeed be used as a biometric. Surprisingly, research into the psychology of gait has
not received much attention, especially using video, in contrast with the enormous
attention paid to face recognition. One recent study [45], using video rather than point
light displays, has shown that humans can indeed recognize people by their gait, and
can learn their gait for purposes of recognition. The study concentrated on
determining whether illumination or length of exposure could impair the ability of
gait perception. The study confirmed that, even under adverse conditions, gait could
still be used as a cue to identity.
Clearly, psychological studies confirm Shakespeare’s earlier observations, and
support the view that gait can indeed be used for recognition. Prior to study of
automatic recognition, we shall consider some of the (many) approaches to human
body and motion modeling, for these are of potential benefit in recognition. Indeed,
some of the approaches have found deployment in automatic gait recognition.
Modeling the Human Body and its Motion
Many studies have considered human motion extraction and tracking, though not for
recognition purposes. The selection of good body models is important to efficiently
recognize human shapes from images and analyze human motion properly. Stick
figure models and volumetric models are commonly used for three-dimensional
tracking, and the ribbon model and blob model are also used but are not so popular.
Stick figure models connect sticks at joints to represent the human body. Akita [1]
proposed a model consisting of six segments: two arms, two legs, the torso and the
head. Lee and Chen’s model [27] uses 14 joints and 17 segments. Guo et al. [18]
represent the human body structure in the silhouette by a stick figure model which has
ten sticks articulated with six joints.
On the other hand, volumetric models are used for a better representation of the
human body. One model [38] consists of 24 segments and 25 joints and those
segments and joints are linked together into a tree-structured skeleton. The “flesh” of
each segment is defined by a collection of spheres located at fixed positions within the
segment’s co-ordinate system. Concurrently, angle limits and collision detection are
incorporated in the motion restrictions of the human model. Among the different
volumetric models, generalized cones are the most commonly used. A generalized
6 Nixon et al.
cone [29] is the surface swept out by moving a cross-section of constant shape but
smoothly varying size along an axis. Generalized cylinders are the simplified case of
generalized cones that have a cross-section of constant shape and size. Fig. 11.3
shows examples for a stick figure model, a cylinder model and a blob model.
stick figure model cylinder model blob model
Figure 11.3 Human body models.
Later work developed Marr’s approach [19, 42] to a set of 14 elliptical cylinders
representing the feet, legs, thighs, hands, arms, upper-arms, head and torso. Kurakake
and Nevatia [26] treat the human body as an articulated object having parts that can be
considered as almost rigid and connected through articulations. They use the ribbon
which is the two-dimensional version of the generalized cylinder to represent the
parts. The blob model was developed by Kauth et al. [24] for application to multispectral
satellite (MSS) imagery and used in human motion tracking [3]. The person is
modeled as a connected set of blobs, each of which serves as one class. Each blob has
a spatial and color Gaussian distribution, and a support map that indicates which
pixels are members of the blobs.
However, these structural models need to be modified according to different
applications and are mainly used in human motion tracking. The alternative is to
consider the property of the spatio-temporal pattern as a whole. Among the current
research, human motion can be defined by the different gestures of body motion,
different athletic sports (tennis, ballet) or human walking or running. The analysis
varies according to different motions. There are two main methods to model human
motion. The first is model-based: after the human body model is selected, the 3-D
structure of the model is recovered from image sequences with [27,41] or without
moving light displays [1,18,19,42]. The second emphasizes determining features of
motion fields without structural reconstruction [28,33,39].
Ideas from human motion studies [34] can be used for modeling the movement of
human walking. Hogg [19] and Rohr [43] use flexion/extension curves for the hip,
knee, shoulder and elbow joints in their walking models. Guo et al. [18] use joint
angles between different sticks as features of different walking persons. A different
approach for the modeling of motion was taken by Akita [1], who used a sequence of
stick figures, called the key frame sequence, to model rough movements of the body.
In his key frame sequence of stick figures, each figure represents a different phase of
body posture from the point view of occlusion. The key frame sequence is determined