You are on page 1of 12

Gait Analysis Review

David C. Post

28 February 2006

Abstract
Gait anaylsis is a a subject that has interested researchers for many years. Recently, it
has been examined for the purposes of biometrics and rehabilitation. This review describes
characteristic gait analysis tools and research in those areas.

1 Introduction
Imagine that you are heading home after a long day at work. As you walk down the hallway
of your apartment complex, groceries in hand, a surveillance camera watches you closely.
As you near your apartment door, you hear a beep and the door opens in front of you.
“Welcome home,” says a computer-generated voice. This sounds like a scene from a a
futuristic movie, but with the advancements in gait recognition it can become a reality. In
our everyday lives we recognize the gait of certain individuals. We quickly notice if someone
is running rather that walking. Additionally, one is able to pick out a friend, just by the
way he or she walks, without see his or her face. For us this is relatively easy. However,
teaching a computer system to recognize gait is a real challenge.
Gait analysis and recognition is an active interest of many computer vision research
groups. Although current interest concerns computer vision, gait has been studied from
several different viewpoints in the past. Those in areas of kinesiology, physiotherapy, or-
thopedic surgery, and ergonomics, have all had interest in gait analysis [2]. Some of their
results are being interpreted in new ways for the computer vision community. According
to BenAbdelkader a biomechanics interpretation of human gait is the “synchronized, in-
tegrated movements of hundreds of muscles and joints in the body.[2]” All humans follow
the same basic walking pattern, but their gaits are influence by functions of their entire
musculo-skeletal structure. Limb lengths, body mass and shape, stride length, and several
other factors influence how a person walks. Given that we as humans are generally unique,
this structure is difficult to duplicate leading to the belief that a few hundred kinematic
parameters can uniquely identify an individual [2]. However, obtaining such a complete
profile from low-resolution video remains an open problem in the research area.
Applications for gait recognition seem to fall into two or three main categories. Gait
recognition as a biometric, and as a tool in rehabilitation or sports activity. As a biometric
gait can identify suspicious individuals or grant entry to those authorized to restricted access
locations. From a rehabilitation standpoint, gait analysis can be used to identify musculo-
skeletal deficiencies and develop programs for such. In the sports world, gait recognition can
be used for training and analysis and improvement of performance. This review presented
characteristic systems for both biometric and rehabiliation applications.

1
2 Biometrics
A biometric is something that can be used to verify the identity of an individual. The most
common biometric that comes to mind is a finger print. Recent events have brought national
interest in quick identification of suspicious individuals. Areas such as airports, parking lots,
banks, and bus/subway stations, all have a need for quick detection of threats. However
current biometrics–fingerprints, and facial recognition, are limited and time consuming.
Trying to fingerprint everyone that walks through an airport is not possible. It is probably
not even legal. A major advantage of gait recognition is that is it unobtrusive. It can
be measured at a distance, without the knowledge or cooperation of the subject, whereas
current methods would require physical touch or close-range sensors [2].
Most gait recognition systems work in the same general way as shown in Figure 1. First
data must be collected from the individual in question. In this step it helps to have the
background be as simple as possible to provide the highest level of recognition. Additionally
selection of an appropriate viewpoint, one in which the gait is observed from the side, is
also important. From here, through a process called background subtraction, the object or
gait is separated from the background noise. Next the specific markers of the identification
scheme are extracted from the gait data. These are compared with the database in hopes
of a positive recognition.

Figure 1: General diagram of gait recognition system[11]

2.1 Model-Based
Methods of identification can be divided into two categories: model-based, and motion-
based. Model-based approached build up a model of how one would walk and then try
to fit their image data to the model. One model takes measurements of static body pa-
rameters, such as the height, distance between head and pelvis, distance between the feet,
and maximum distance between the pelvis and the feet(Figure 2(a)). In another model the
silhouette was divided into seven elliptical regions and properties of these ellipses studied
over time (Figure 2(b)). Another model proposed measuring the angle of hip rotation, while
another method used different shapes to characterize the silhouette (Figure 2(c),(d)).
Figure 2: Model-based approaches [11].

2.2 Motion-based
Motion-based approaches are sort of the opposite, capturing the outline of a motion and
trying to analyze it, based on data already collected. The image is though of as just a
shape, not having human parameters associated with it. Many motion-based approaches
use silhouettes of the gait. Examples of these silhouettes were as shown in Figure 3. These
binary images are good to use because they are computationally simple, not relying on color
or texture data that may not be available from the motion capture. Some approaches convert
these images to 1D or 2D signals. For example, one method might extract the silhouette
boundaries at different times during the gait and compare them to those signals developed
from the database [2]. It is a process akin to a Fourier Transform. Other approaches use
the image as is, and attempt to identify by comparing to a gallery of existing images.

Figure 3: Example Silhouettes [9].

Due to differences in camera angles and distances, methods have been developed to
scale the shapes. One example of such is Procrustes shape analysis. It consists of defining
the centroid of two dimensional shape and defining the boundary of the shape as a vector
of elements,
z = [z1 , z2 , ..., zn ]T , (1)
where zn = xn + i ∗ yn a complex number describing a point with x i and yi components.
Figure 4 showed this graphically. We may consider two shapes, z 1 and z2 , equal if under a
combination of translations, scalings, and rotations,

z1 = α1k + βz2 , α, β ∈ C (2)


j6 β
β = |β|e (3)

where α1k translates z2 and |β| and 6 β scale and rotate z2 , then they have the same
configuration [10].
Figure 4: Boundary of a Silhouette [10].

Since these silhouettes are the basis for identification, it is important that they be
as clear and noise free as possible. Looking back at Figure 3 one can see that the image
boundaries are rather choppy. A research group from the University of South Florida is
working on a method to improve these silhouettes. When the silhouettes are produced
by eliminating the background several errors can occur. Shadows, moving objects in the
background, and compression artifacts from the capture devices are just a few of these. To
eliminate these, frames are first identified from characteristic points in the gait, such as,
both feet down-right leg forward, initial step, swing, landing, and so on. These frames are
picked and grouped by computer and then averaged together over all subjects to create
what they call eigen-stances. Representations of these can be seen in Figure 5. Then all the
data is compared with its characteristic eigen-stance and an algorithm they have developed
reconstructs or clean-ups the images. Results of their method can be seen in Figure 6
[7]. Figure 7 obtained from [10] shows further proof of the uniqueness of gait. Part (a)
of Figure 7 shows one subject’s gait outlines compared to his exemplar or eigen-stance for
that point. Notice how they barely differ. Part (b) shows the contours of several different
subjects at the same point in their gait. Notice the wide variability. This further suggests
the uniqueness of gait to an individual.

Figure 5: Eigen-Stances [7].

To use this method of comparing shapes to a database or gallery, one first needs to
construct a gallery. One such database has already been constructed for the HumanID
Gait Challenge Problem. Data consists of 1870 gait sequences from 122 individuals. These
sequences were collect by varying five different parameters: two camera angles, two shoe
types, two walking surfaces, with or without carrying a briefcase, and two different times
of year [7]. This data is available to those who are interested in developing algorithms to
analyze it and many research teams are centering their work around this common data
set. Additionally this project is being sponsored by the U.S. Defense Advanced Research
Project Agency (DARPA) [11].
Figure 6: Reconstructed Silhouettes [7].

Figure 7: Gaits compared to Eigen-stance and variability of eigen-stances between individ-


uals [10].

2.3 Limitations
However, there are limitations to gait analysis as a sole biometric identifier. Unlike other
identifiers, notably fingerprints, gait is prone to a wide variability within a specific subject.
Research shows that several factors can vary the gait of an individual: shoes, clothing,
terrain, mood, illness, fatigue, and others. However, there is evidence from the original
psychophysical experiments that suggests that a gait signature is possibly unique to each
individual [2]. The literature suggest that as of yet, gait cannot be used by itself, but
combine with other identifiers it yields substantial results [11].

3 Rehabilitation
When gait analysis is used as part of rehabilitation treatment or performance enhancement,
several factors change. First and formost, the method is no longer unobtrusive. The subjects
have agreed to take part in the gait anaylsis and thus more elaborate and percise collection
systems can be used. Whole labs are dedicated to data collection, and much hardware and
software is available. When taking advantage of availble resources much can be studied
about gait and applied to rehabilitation.
3.1 Gait Capture Systems
With the subject’s knowledge and willingness to participate, a more elaborate setup can be
used to track their gait. A conjunction of several different systems can be used to obtain
vast amounts of data about a specific gait trial. Cameras, magnetic fields, electromyography
(EMG) data, and force plates are all methods used for data collection
One method uses sensors placed on the body and a transmitter which generates a
low-frequency magnetic field. These are connected to a control unit that allows computer
software to track the positions and rotations of these sensors in 3D space. Inverse kinematics
are used to solve for the various joint angles, taking into account the sensor offsets from
actual joint centers. However, this method is very sensitive to metal in the environment
structure (floors, walls, ceilings) and other environmental magnetic fields, such as those
from near by electrical equpiment. Ultrasound waves can be used in much the same way as
the magnetic field.
Another approach mounts cameras around the room, each of which has an infra-red-
emitting diode on it. Reflective markers are placed in the same way as in the magnetic
approach. The Helen Hayes pattern is a common one used for marker placement and is
shown in Figure 8. The marker arangement allows both joint angles and body segment
rotations to be observed. The markers reflect the infra-red light, which is picked up by
cameras in a direct line of sight to the marker. A typical camera setup was shown in
Figure 9. As long as two cameras can see the marker, its position can be calculated from
the calibrated volume of the camera space. This method is nice because the markers are
completely wireless. However several errors develop with this method, including swapping
markers, noisy data, markers falling off, and false reflections. However, comparing data to
a human model, geting a relative idea of the most probably marker positions, can cut down
on extraneous data.

Figure 8: Helen Hayes Marker Locations [6]

To collect EMG data, one method places electrodes over the gluteus medius, vastus
lateralis, and medial gastrocnemius. These are the major muscles used in a normal walking
stride. Activation magnitudes are then normalized to levels during maximal effor manual
Figure 9: Typical Capture System [8]

muscle testing (MMT) [5]. Finally, some setups have force-plates in the floor to help collect
that data as well.

3.2 Categorizing Gait using a ANN


Hahn and Chou developed a model for estimating the falls risk in elderly people. Many
elderly people are prone to falls because of reduced muscle strength and sense of balance.
Over 35 percent of the elderly population experiences falls with approximately 20.2 billion
dollars spent on treatement each year [4]. Data was collected from 29 elderly subjects.
These subjects consisted of 19 healthy individuals and 10 “fallers.” A “faller” was one
with a self-reported history of two or more falls. Healthy subjects were subject to a list
of critera as listed in [4]. Data was averaged within each individual over three indiviudal
gait trials. Data taken consisted of EMG data; gait velocity, stride length, stride time,
step width (temporal-distance (T-D)); medio-lateral(M-L) motion (displacement and peak
velocity) of the body center of mass (COM). The body motion data (all but EMG) was
collected using a six-camera ExpertVision system produced by the Motion Analysis Corp.
A similar system is as shown back in Figure 9. The trial consisted of walking on level ground
with no obstructions then stepping over an obstacle equal to 2.5% percent of a subject’s
height. The trial was performed at the subject’s own pace, while barefoot.
Various combinations of the data were used as inputs to an artificial neural network
(ANN). The goal of the ANN was to catergorize input as faller(1) or non-faller(0). A three
layer feed-forward network was used. Training was done with 70% of the sample data.
Several configurations of inputs, hidden layer neurons, and error goals were tested. These
were as shown in Figure 10. The best results were for a hidden layer with five neurons,
error goal of E = 0.001, and inputs of EMG and T-D data. A diagram of the ANN was as
shown in Figure 11. A sigmoidal function was used in the hidden units and a pure linear
function used on the output unit. The network had one single output neuron which was
used to categorize the individual.
After training was completed, cases from the testing set, the remaining 30% of the data
was input into the network with the fixed, learned biases and weights. From the output of
the network a distribution function was generated. Correct categorization required picking
a cut-off line for predicting the data was that of a faller. A conservative decisioon line X 0
was picked at the median of the healthy patient’s output disrtibution. Each training session
calculated a new X0 . An output distribution was also developed for the faller data, in order
to test the validity of the decision line selection. A graphic of these distributions was shown
Figure 10: ANN Configurations [4]

in Figure 12.
The effectiveness of categorization was measured by the relative operating characteristic
value (ROC). It is a method from signal detection theory which is supposed to provide “a
precise and valid measure of diagnostic accuracy[4].” The ROC value is the area under
the curve of specificity versus sensitivity for multiple testing sessions. Specificity is the
proportion of false-positives, while sensitivity is the proportion of true positives. For the
configuration give earlier their ROC was the highest with ROC = 0.890. This method is
widely used in medical diagnostics[4].
A metric for risk estimation was developed. The output value of the testing case, X 1
was used to develop,
1 − X1
Dr = , (4)
1 − X0
where Dr is the relative distance of the predicted value from X 0 . As such, when
the metric aprroached or exceeds 1.0 the relative risk of falling decreases. As the value
approaches or exceeds 0.0 the risk of falling increases. For example, a value of -1.00 indicated
very high risk[4]. Their method of using a network to predict fall risk is in contrast to other
methods that rely more on evaluation of the subject’s visual, cognitive, and depressive
states.

3.3 A gait rehabilitation system


Several groups of people need gait rehabilitation. Sports injuries (ACL) and stroke victims
are just two categories that readily come to mind. During treatment for gait rehabilitation,
physical therapists use words and physically guiding the patient’s body to explain the motion
for walking. Thsi approach is problematics because, there’s a shortage of therapists for
those requiring treatment. Additionally, the therapist could quickly tire from the repetitive
motion. Finally, even in their expertise it is still difficult for a therapist to teach natural-
looking leg movements to patientsc̃iteYano. As such robotic systems are being developed
to alleiviate this burden.
One such interface, called the GaitMaster2 (GM2), consists of “two 2-DOF platforms
, which are chain-drive jacks equipped with an AC servomotore and an optical rotary
encoder[12].” It is very similar to an elliptical stepper machine found in most gyms. A
photo and cartoon of the apparatus were shown in Figure 13. They performed several gait
training trials with hemiplegic patients. They took video and EMG data of the patient’s
Figure 11: ANN Diagram [4]

Figure 12: ANN output distribution curves [4]

gait cycle before and after training on the GM2. One trial concerned a stoke patient w/
a paralytic left-leg. After the 3 month trial the patient was able to keep her balance and
transfer her weight to her paralytic leg. This was a task she was unable to perform before
and thus her quality of life was greatly improvedc̃iteYano.
The system in and of itself is not neccessarily very intelligent. It just takes the input
data and produces the pre-determined motion path. However, the intelligence comes from
where that data is obtained and interpreted. For example, one could motion capture then
individual’s unassisted gait. Following the ideas from the elderly fall section, its foreseeable
that one could teach a ANN to distingush between different degenearacies in the unassisted
gait. Gioftsos and Grieve developed an experiment to test a similar idea. Gait data was
taken from a group of 10 males and 10 females for 3 gait patterns. The first pattern was
Figure 13: The GaitMaster2 system [12]

normal walking. The second required walking with a 3.5 kg mass securely strapped to the
right ankle. The third used a knee brace to lock the right knee in an extended position[3].
The second and third gait patterns are thought to mimic those of a hemiplegic patient. Their
test confirmed that a neural network was able to identify the gait pattern with acceptable
accuracy[3].
Given this information, the author believes a system could be designed similar to the
following. Data is first collected from the patient’s unassisted hemiplegic gait. This data is
fed to the trained ANN to classify its pattern. This pattern classification can be used to
select the most appropriate rehabilitation gait pattern from a library of collected patterns.
These collected patterns would be categorized by another system for their usefullness in
rehabilitation the proposed gait. Once the most appropriate pattern is found it is scaled to
the specific patient and input to the GM2 for rehabilitation training. As such, a very low
intelligence piece of equipment can become part of a highly intelligent system.

3.4 Leaving the Lab


Motion captures systems as have been described do have their limitations. As such Aminian
and Najafi describe several devices that could be useful as body-fixed sensors for on-the-
go data collection. They suggest that body-fixed sensors allow outdoor and mobile mtion
caputre. The sampling frequency for the on-the-go systems is also much higher. Finally
since there are no markers, signals can be continuously recorded without and loss due to
marker problems discussed earlier.
Several of these sensors are described below. A goinometer is a device which can
measure the angular displacement of a joint. It consists of two attachement bars with flexible
composite cylinder in between. Mounted to the cylinder are several strain gauges, which
were used to relate a voltage to an angle. Foot pressure sensors are also being developed.
These aren’t accurate enough to measure force, but rather are used to measure the period of
different phases of normal walking (heel strike, swing and stance phase, heel and toe off[1]).
Accelerometers are another suggested tool, and by using two or more of these attached to
different body segments, resting body posture was recognized. Even with the influence of
gravational effects, Aminian and Najafi previously reported sucess in clinical applications
of the devices. Using Micro-electromechanical systems (MEMS) gyroscopes for human
motion capture is still under development. However, they are currently used in automotive
applications and anti-jitter compensators for video-cameras. The MEMS gyroscopes do not
have the influence of gravity in their signal like the accelerometers do. Studies have shown
that the estimation of knee angle and trunk tilt are possible. However, gyroscopes have
drawbacks such as power consumption, prive, drift, and sensitivity to shock[1].
They propose a sensor composed of two acceleromters and a gyroscope. The accelerom-
eter provided the derivative of angluar velocity and an inclination angle at rest, while a gy-
rocsope estimates the derivative of an angle. Together they can estimate the 2-D kinematics
of a body segment. Using these sensors and a lightweight data logger (300g) the subject
can be set free to an natural outdoor environment to perform natural motion. Figure 14
shows an example of their system.
Experiments were carried out with sensors monitoring the sagittal-plane only. The data
logger was capable of running for 24 hours without needing a recharge. In their tests they
were able to distinguish differences in posture from sitting, standing, lying, walking, and
jumping. They also were able to collect data from hip replacement surgery patients at 3, 6,
and 9 month intervals before and after surgery. This data proved to be much more valuable
in the rehabilitation process than standard doctor-patient inquiries and evaluations [1].

Figure 14: Portable data collection configuration[1]

4 Conclusions
Gait analysis has been the subject of research for several years. Most recently those in
the computer vision and security fiels have begun to analyze gait for use as a biometric.
Although current results are promising, the implementation of such a system is still some
time away. Gait anaylsis has been used in some form or another in rehabilitation from
the very beginning. Newer and better systems are being developed to take advantage of
itelligent analysis tools (ANN, etc.) that have recently been applied to the area. The future
in these fields is both promising and exciting. It won’t be long before these systems can be
applied to improve our everyday lives.
References
[1] Kamiar Aminian and Bijan Najafi. Capturing human motion using body-fixed sensors:
Outdoor measurement and clinical applications. Computer Animation and Virtual
Worlds, 15(2):79 – 94, 2004.

[2] Chiraz BenAbdelkader, Ross G. Cutler, and Larry S. Davls. Gait recognition using
image self-similarity. Eurasip Journal on Applied Signal Processing, 2004(4):572 – 585,
2004.

[3] G. Gioftsos and D.W. Grieve. Use of neural networks to recognize patterns of human
movement: gait patterns. Clinical Biomechanics, 10(4):179 – 183, 1995.

[4] Michael E. Hahn and Li-Shan Chou. A model for detecting balance impairment and
estimating falls risk in the elderly. Annals of Biomedical Engineering, 33(6):811 – 820,
2005.

[5] Michael E. Hahn, Arthur M. Farley, Victor Lin, and Li-Shan Chou. Neural network
estimation of balance control during locomotion. Journal of Biomechanics, 38(4):717
– 724, 2005.

[6] http://www.lifemodeler.com. Helen hayes markers graphic.

[7] Zongyi Liu and Sudeep Sarkar. Effect of silhouette quality on hard problems in gait
recognition. IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernet-
ics, 35(2):170 – 183, 2005.

[8] J.C.T.B. Moraes, S.W.S. Silva, and L.R. Battistela. Comparison of two software pack-
ages for data analysis at gait laboratories. Annual International Conference of the
IEEE Engineering in Medicine and Biology - Proceedings, 2:1780 – 1783, 2003.

[9] P.J. Phillips, S. Sarkar, I. Robledo, P. Grother, and K. Bowyer. Baseline results for
the challenge problem of humanid using gait analysis. Proceedings of the Fifth IEEE
International Conference on Automatic Face and Gesture Recognition, pages 130–135,
2002.

[10] Liang Wang, Huazhong Ning, Tieniu Tan, and Weiming Hu. Fusion of static and
dynamic body biometrics for gait recognition. Proceedings of the IEEE International
Conference on Computer Vision, 2:1449 – 1454, 2003.

[11] Bruce J. West and Nicola Scafetta. Gait recognition: A challenging signal processing
technology for biometric identification. IEEE Signal Processing Magazine, 67(5 1):78–
90, 2005.

[12] Hiroaki Yano, Kaori Kasai, Hideyuki Saitou, and Hiroo Iwata. Development of a
gait rehabilitation system using a locomotion interface. Journal of Visualization and
Computer Animation, 14(5):243 – 252, 2003.

You might also like