Evaluation of KINECT and SHIMMER Sensors for Detection of Gait
Parameters
Katja Orlowski and Harald Loose
Department of Computer Science and Media, Brandenburg University of Applied Sciences,
Magdeburger Str. 50, 14770 Brandenburg, Germany
Keywords:
Motion and Gait Analysis, KINECT and SHIMMER Sensors, Low-cost Sensors, Health Applications.
Abstract:
Detecting gait parameters is possible using various sensors based on different physical principles. In our
investigation a visual system, the Microsoft KINECT, and an inertial sensor system with SHIMMER 9DoF-
sensors, are used for capturing the gait of various persons. Both systems have a small form factor and are
affordable regarding cost. Hence these are well-suited for mobile applications in the health care environment.
Using these low-cost sensor systems, motion capture and analysis can be done in hospitals, physiotherapy units
or nursing homes. This paper focusses on the comparison of detected gait parameters by analyzing statistical
parameters. The examination of accuracy of both systems is carried out in two steps; first by initially measuring
the gait of a small group of volunteers and second of a larger group. The noise is also examined which has to
be filtered out in the preprocessing procedure. The choice of filter impacts the detection of gait parameters.
As a result the noise is characterized rather nonspecifically in both systems. As expected, the gait parameters
determined by the systems are not identical, but similar. The deviations vary in the specific gait parameters;
some are less error-prone than other.
1 INTRODUCTION
Nowadays smart, mobile sensors are applicable in
health care. In previous papers their applicability was
discussed and first results were shown (see (Orlowski
et al., 2012)). In this paper the algorithms used for
the automatic detection of gait parameters are intro-
duced based on the analysis of the normed gait cycle
by (Perry, 2010). Perry calculated the mean joint an-
gles of the foot, knee, thigh and hip from the gait of
55 subjects. Derived values of those data are used to
check the correctness of the algorithms. The given
normed gait cycle consists of a stance and a swing
phase. The stance starts with putting the heel on the
ground (initial contact, IC) and ends with the detach-
ment of the toe from the floor (terminal contact, TC).
The TC is the beginning of the following swing phase
which ends with IC. While one foot is in the swing
phase, the other has full contact to the floor. Perry
et al. subdivided the stance in five, and the swing in
three phases. In our investigation only three of eight
phases are considered. In addition to stance and swing
the midswing, the middle third of the swing phase, is
investigated. The midswing begins when the swing-
ing foot crosses the standing one and ends when the
Figure 1: Gait cycle - left (black) and right (white) leg with
the swing and stance phase marked by colored stripes and
IC, TC, midswing and -stance (Murray and Kory, 1964).
”swinging limb is forward and the tibia is vertical”
(Perry, 2010). It is the interval from about 75-87 % of
the gait cycle.
Figure 1 shows a whole stride for each leg (left
- black, right - white) and the percentage of walking
cycle. An overlap of the stance phases of both legs is
identifiable. During normal walking, the swing phase
is always shorter than the stance, and the swing phase
of one leg is within the stance phase of the other. A
40 to 60 percent relationship between swing to stance
phase is generally assumed (Perry, 2010).
This paper includes a comparison of the gait pa-
rameters of the Kinect and SHIMMER sensors based
on the data of a step sequence of 26 volunteers. Fur-
thermore, the sensor noise is investigated through sep-
arate experiments as the choice of the filter influences
157
Orlowski K. and Loose H..
Evaluation of KINECT and SHIMMER Sensors for Detection of Gait Parameters.
DOI: 10.5220/0004227901570162
In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 157-162
ISBN: 978-989-8565-36-5
Copyright
c
2013 SCITEPRESS (Science and Technology Publications, Lda.)
the quality of the results. The accuracy of the sensors
is evaluated using statistical values.
2 SYSTEMS AND EXPERIMENTS
2.1 Systems
For motion capture and analysis different sensor types
are used, which can generally be distinguished due
to different recording principles and fields of ap-
plication. The Microsoft KINECT sensor system
(KINECT) and the Shimmer
TM
sensors (SHIMMER)
are predestined for mobile health applications be-
cause of their small size (see Fig. 2) and the compar-
atively low cost. While KINECT is a visual system
based on an absolute reference system, the SHIM-
MER (9-DoF-sensor) is an inertial system working
with a relative reference system. That system cal-
culates incrementally the position of the sensor by
knowing the initial position.
The depth image of the KINECT is produced by
measuring the distortion of the reflected dots from the
pseudo-random beam pattern sprayed out from the
IR laser (Taylor, 2011). Further information on the
used systems can be found in earlier papers (Orlowski
et al., 2012; Orlowski and Loose, 2012)
2.2 Experiments
Preliminary investigations were done to determine the
noise of the sensor systems. For this experiment a
dummy (see figure 2) was used. In contrast to a hu-
man test person moving slightly during standing, a
skeleton always rests. The KINECT is able to detect
correctly the skeleton and to capture the stands. For
measuring the SHIMMER sensor noise, the calibrated
sensor is placed on a table. The results of these anal-
yses are describe in the following subsection.
To assess the determination of gait parameters us-
ing KINECT and SHIMMER sensors, two steps have
been carried out: first the gait of a small group of vol-
unteers is measured for an initial assessment of the
setup and algorithms, thereafter the data of a larger
group of healthy subjects is evaluated.
Within the small group the step sequence of six
healthy subjects (3 M, 3 F, mean age: 29.8) is si-
multaneously captured with both systems. Due to the
restricted capturing area of the KINECT the step se-
quence consists of one to three steps each leg, vary-
ing in step length. The KINECT is placed in front
of the walking line at a defined height. The person
walks toward the KINECT until a distance of 80 cm is
reached. During the walking experiment the proband
Figure 2: The Microsoft KINECT and two SHIMMER sen-
sors (left). The unmoved dummy used to determine the
KINECT sensor noise (right). Below: Experimental setup
for noise evaluation with marked dummy positions.
wears two SHIMMER 9-DoF-sensors which are fixed
above the ankle on each leg. A start and stop synchro-
nization is not implemented; the systems are started
and stopped manually. After the initial assessment
the gait of 26 healthy students (20 M, 6 F, mean age:
24.1)) was captured. There were no special in- or ex-
clusion criteria for the choice of the volunteers.
2.3 Noise Evaluation
2.3.1 KINECT
In contrast to the publication of (Khoshelham and El-
berink, 2012), the sensor noise of the KINECT is
reviewed by measuring the skeletal data of the non-
moving dummy. To verify the hypothesis that the
noise is dependent on the distance, the skeleton was
recorded at five positions (see figure 2). To elimi-
nate random noise of a single KINECT, the setup was
repeated with two other KINECT systems. Three 30-
second-datasets at each given distance with a frame
rate of 30 Hz were recorded.
The noise signal of selected joints and their fre-
quency spectrums were examined. Most of the fre-
quencies of the noise signal are below 5 Hz with a
more or less normal distribution. The reason for the
slightly higher proportion of low frequencies might
be a drift. The mean, standard deviation (std), mini-
mum and maximum are determined as statistical val-
ues. The mean of the standard deviation of the hip is
±1.3 mm, of the shoulder ±2.7 mm and of both ankles
±3.4 mm.
Figure 3 shows the box plots of the nine record-
ings. The samples of the selected joints of the nine
recordings are each summed up to one vector consist-
ing of about 8100 samples. The median value is close
to zero because the data is shifted by its mean. Obvi-
ously, the deviations of the left and right side are sim-
BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
158
Figure 3: Box plots: noise of both ankle, hip and shoulder
summarizing the measurements of 3 KINECT (dist 3 m).
Table 1: Mean of the standard deviation of the noise and its
relative value depending on the distance.
Dist
[mm]
left
Ankle
(relErr)
right
Ankle
(relErr)
Hip
(relErr)
Shoulder
(relErr)
2300 1,5 mm
(0,07 %)
2,3 mm
(0,1 %)
0,9 mm
(0,04 %)
1,4 mm
(0,06 %)
3000 3,1 mm
(0,10 %)
3,4 mm
(0,11 %)
1,3 mm
(0,04 %)
2,7 mm
(0,09 %)
4000 6,1 mm
(0,15 %)
10,5 mm
(0,26 %)
3,4 mm
(0,09 %)
5,0 mm
(0,12 %)
ilar with the exception of more outliers on the right
side. The center of the shoulder has larger deviations
than the hip center. The deviations including the out-
liers of the hip and shoulder are much smaller than
those of both ankles.
Further evaluations compare the dependency of
the noise on different distances. Even when using the
smoothing parameter of the KINECT SDK for less
skeletal jitter (see Microsoft KINECT SDK API Ref-
erence or (Fernandez, 2011)), the noise grows with
the distance. Table 1 contains the calculated mean of
the standard deviation of the recordings at different
distances as well as the relative errors based on the
respective distances.
2.3.2 SHIMMER
The noise of four SHIMMER sensors was deter-
mined. After sensor calibration a 25-second record-
ing was performed. The signals of the gyrometer were
transformed to the frequency domain. Examining the
recorded noise of the gyrometer, the noise could be
classified rather nonspecific. It is more or less nor-
mal distributed and has a few distinct peaks. Similar
Figure 4: Box plots of the noise of the gyrometer.
to the KINECT noise evaluation, statistical values are
determined for all noise signals. Figure 4 represents
a summary of the statistical evaluation of the noise
signals. All angular velocity signals in x, y and z are
summed up to one vector each, and box plots were
created. The plot contains the box plots of the angu-
lar velocity. Obviously, there are no significant differ-
ences between the three axes. The standard deviation
of the angular velocity is ±0.3
/s .
3 METHODS
3.1 Normed Gait Cycle
Using the mean joint angles provided by Perry et
al., different data, such as angular velocity, horizon-
tal movement and velocity, are derived from the mean
joint angles of the left leg (see figure 5) using the limb
length of a subject (ankle-hip: 88 cm, ankle-knee: 41
cm, knee-hip: 47 cm, foot: 28 cm) . Since the speed
of a normal gait is defined at 1.3-1.6 m/s (Oeberg and
Oeberg, 1993), we choose an uniform forward motion
of the hip at 1.4 m/s (see figure 6 (below)). The values
of derived velocity are negative because the distance
to the sensor is reduced during the measurement.
Figure 5: Mean joint angles of the normed gait cycle.
All diagrams in figure 6 as well as the walking cy-
cle in figure 1 show that the stance phase of the left
leg is approximately between 5 and 60 % of the gait
cylce. The swing phase starts directly after the stance
phase and is not yet completed at 100 %. The angu-
lar and the horizontal velocity are around zero dur-
ing stance since the distance to the sensor does not
change. The middle diagram, depicting the horizon-
tal movement of the foot, ankle, knee and hip, shows
EvaluationofKINECTandSHIMMERSensorsforDetectionofGaitParameters
159
Figure 6: Calculated angular velocity (above) of the shank
and foot, horizontal velocity (middle) of foot, ankle and
knee and horizontal movement (below)
a nearly constant distance around 3 m from 5 to 50-60
% of the gait cycle. After analyzing the normed gait
and the derived values, conclusions for the determi-
nation of gait parameters can be drawn. It is possible
to convey characteristics of the gait which can be ex-
tracted automatically from the measured data of the
KINECT and SHIMMER gyroscope.
A gait cycle consists of a stance and a swing phase
and can be characterized by its length and duration.
From the angular velocity of the shank, the begin-
ning (IC) and end (TC) of the stance phase as well
as the middle of swing phase (midswing) can be de-
termined. In contrast to the definition of midswing
as a phase (Perry, 2010), midswing is used as a char-
acteristic point of the swing phase. At the beginning
and end of the stance phase local minima occur in the
angular velocity. The midswing is characterised by a
local maxima (see figure 7).
The beginning and end of the swing phase can be
detected in the velocity of the foot motion using a
threshold. During swing phase this threshold is ex-
ceeed. The beginning and end of the stance result in
the preceding swing phase parameters. In figure 7 the
lime, solid curve represents the velocity of the foot,
the dashed line the used threshold. The red circles
show where the threshold is exceeded.
Figure 7: Velocity and angular velocity with marked param-
eters and used threshold.
3.2 Determination of Gait Parameters
3.2.1 SHIMMER
The swing and stance phases are determined by iden-
tifying the characteristic points from the data of the
gyroscope. Figure 8 shows signals captured during
normal walking. The angular velocity captured with
SHIMMER is similar to those of the normed gait cy-
cle (compare figure 7 and figure 8).
The IC, TC and midswing points are determined
using an adapted algorithm analogical to that pro-
posed by Greene et al. (Greene, 2010). First dis-
turbances are minimized by a low-pass filter (5th
order Butterworth filter, corner frequency: 5 Hz).
Then the correctness of the polarity of the signal has
to be checked using the formula for the skewness
(mean(x) median(x)). If the skew value is nega-
tive and the absolute value of the minimum is greater
than the absolute value of the maximum, the signal
is mirrored at the x axis. An additional criterion
(abs(min(x)) > abs(max(x))) is used. The combi-
nation of both criteria can change the polarity of the
signal. The local maxima and minima are searched
within the signal. Each local maximum stands for the
midswing of a swing phase. The preceding local min-
imum of the midswing point is the TC, the succeeding
local minimum represents the IC.
Figure 8: Angular velocity of the left/right lower leg with
marked features: IC-green, TC-yellow, midswing-black.
BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
160
3.2.2 KINECT
The start and end of a swing phase, which correspond
to the terms inital and terminal contact (IC/TC), are
calculated from the ankles‘ forward motion.
The captured data are filtered by a mean filter of
length five which is applied through convolution. The
smoothed signal of the forward motion of the left
(blue-dotted) and right (red) ankle is displayed above
in figure 9. As mentioned before, to determine the
start and end points, the first discrete derivative is cal-
culated. The result of the derivation is shown in the
lower plot in figure 9. At first the local minima are
detected to determine the number of steps for each
leg. That information is necessary for the detection
of start and end points of each swing. We assume
that the local minima in the first derivative represent
the midswing points. An experimentally determined
threshold is each time adapted to the minimum of the
two global minima of the derivative of the right and
left signal (th = 0.499 max) to find all start and
end points of the swing phases.
Figure 9: Forward motion of the both ankles of a step se-
quences of one person: distance and velocity (below)
The upper plot of figure 9 contains the detected
start and end points of the swing phase. The constant
parts represent the stances, the small decreasing parts
the swing phases. The foot as well as the ankle move
only slightly during the stance phase while the move-
ment is clearly visible during the swing phases. The
step sequence begins with the right leg after an initial
recovery (about three seconds). Three right steps and
two left steps have be made during the short walking.
4 RESULTS
4.1 Introductory Remarks
As mentioned before the experiments were done in
two steps. First an initial assessment was performed
with a small group of volunteers to check the systems
and the used algorithms. Then, the gait of a larger
group was captured under the same conditions.
The captured data were visually surveyed to iden-
tify data with measuring failures such as storage er-
rors, early breaks of capturing and data with a dis-
turbed initial recovery. Recordings with such failures
were excluded from the assessment and evaluation.
After removing erroneous data, 55 datasets are in-
vestigated and the characteristic points (IC, TC and
midswing) can be detected.
4.2 Comparison and Statistical
Evaluation
The IC, TC and midswing points were detected us-
ing the described algorithms. The duration of the
swing and stance phase as well as the distance be-
tween midswing points were calculated. The devia-
tions between both systems were measured determin-
ing the absolute error in seconds. The existing devi-
ations are due to the preprocessing steps, algorithms
and the accuracy of the systems.
Figure 10: Box plots of the absolute errors between gait
parameters detected from SHIMMER and KINECT data.
Left: 1-Stance left, 2-Stance right, 3-Swing left, 4-Swing
right, Right: 1-Midswing left, 2-Midswing right
Figure 10 provides information about the devia-
tions of both systems/algorithms during measurement
and shows the absolute error measured for the param-
eters: stance left (1) and right (2), swing left (3) and
right (4) as well as midswing left (1) and right (2).
The absolute error constitutes the difference between
the parameters determined from the data of both sys-
tems. The absolute error of the swings is greater
than those of the stances. The large deviations in the
swing phase result from the fact that the pull of the
foot to the end position of the recording is most of
the time not correctly covered by the KINECT (see
figure 9). Contrary to the SHIMMER, the KINECT
algorithm does not recognize the last swing due to
the used threshold. A reduction of the threshold to
35% of the global minimum leads to the recognition
of the swing, but the duration of that swing differs sig-
nificantly from that determined by SHIMMER. That
could be reasoned by the more difficult recognition
EvaluationofKINECTandSHIMMERSensorsforDetectionofGaitParameters
161
of the last swing. The algorithm has to be optimized
regarding that problem.
The mean absolute error at stance phases is 0.084
s (±0.07 s) (left) and 0.078 s (±0.07 s) (right) as well
as at swing phases are -0.053 s (±0.13 s) (left) and
-0.053 s (±0.12 s) (right). The distance between the
midswing points was also considered. As expected,
this gait parameter is less prone. This is reflected in
the right plot. The mean absolute error for the dis-
tance of midswing points is -0.038 s (±0.04 s) (left)
and -0.042 s (±0.06 s) (right).
5 DISCUSSION
The evaluation of noise provides no clear results that
the signals captured with KINECT and SHIMMER
include a specific type of noise. The noise of the
KINECT depends on the subjects distance to the
recording unit. The greater the distance, the stronger
the influence of noise. Regarding other influences,
the environment, in which the measurements are con-
ducted, has to be investigated in further experiments.
The evaluation of derived values of the mean joint
angles have shown that the signals measured with
KINECT and SHIMMER are usable for the detec-
tion of gait parameters. In general the detected points
are correctly determined. The detected IC, TC and
midswing points represent characteristic points of the
swing and stance. They can be used to characterize
the gait and calculate the length and duration of each
swing or stance. The quality of the detection depends
on the signals’ frame rate and the noise.
Statistical values are used to assess the derived
gait parameters. The absolute errors were determined
by calculating the absolute difference of the gait pa-
rameters determined from the KINECT and SHIM-
MER data. The duration of the swing and stance
phase is not identical but similar with small devia-
tions. The deviations of the swing phase is greater
than those of the stance phases due to the detection
problem of the last swing. The differences between
the distance of the detected midswings are much
smaller than those of the swing and stance phase. The
different frame rates (KINECT: 30 Hz, SHIMMER
51.2 Hz) can be a reason for the deviations because
the temporal resolution is different and the accuracy
is not the same.
6 CONCLUSIONS
This paper presents the influence of noise of both sen-
sor systems. A further component of the paper is
the evalution of the normed gait cycle given by Perry
(Perry, 2010) and Murray (Murray and Kory, 1964).
Based on the evaluation, the correctness of the devel-
oped algorithm was verified. They have been used to
determine the gait parameters duration of swing and
stance phases and distance between midswings. The
investigation was carried out in two steps. At first
the gait of a small group of volunteers was measured
and investigated to check for necessary adaptions on
the algorithms. Consequently, the algorithms were
adapted and additional functions were developed to
ensure a semi-automatic evaluation of amount of data.
Then, in a second step, the gait of 26 healthy students
was recorded and analyzed. The results of the com-
parison are shown.
In further experiments, the comparison with a gold
standard as well as with an established mobile sensor
systems, such as xsens
1
, has to be applied to assess
the quality and correctness of the data. Moreover, the
detection problem of the last swing has to be solved.
REFERENCES
Fernandez, D. (2011). Skeletal tracking fundamentals (beta
2 sdk). http://channel9.msdn.com/Series/KinectSDK
Quickstarts/Skeletal-Tracking-Fundamentals. Online:
16.02.2012.
Greene, B. e. a. (2010). An adaptive gyroscope-based algo-
rithm for temporal gait analysis. Med Biol Eng Com-
put, 48:1251–1260.
Khoshelham, K. and Elberink, O. (2012). Accuracy and
resolution of kinect depth data for indoor mapping
applications. Sensors : journal on the science and
technology of sensors and biosensors : open access,
12:1437–1454.
Murray, M.P.; Drought, A. and Kory, R. (1964). Walk-
ing patterns of normal men. J Bone Joint Surg Am,
46A:335 – 360.
Oeberg, T.; Karsznia, A. and Oeberg, K. (1993). Basic gait
parameters : Reference data for normal subjects, 10-
79 years of age. Journal of Rehabilitation Research
and Development, 30(2):210–223.
Orlowski, K. and Loose, H. (2012). Low-cost locomo-
tion sensors. Proceedings of the 13. Nachwuchswis-
senschaftlerkonferenz 2012, 13.
Orlowski, K., Loose, H., Otte, K., Mansow-Model, S., and
Thiers, A. (2012). Kinect and shimmer sensors in mo-
tion analysis in health applications. Proceedings of
BIOSIGNALS 2012, Int. Conference on Bio-Inspired
Systems and Signal Processing, Vilamoura, Algarve,
Portugal, 1-4 Feb. 2012, pages 226–231.
Perry, J. (2010). Gait Analysis - Normal and Pathological
Function. Slack Inc.
Taylor, T. (2011). Kinect for robotics. http://blogs.
msdn.com/b/msroboticsstudio/archive/2011/11/29/
kinect-for-robotics.aspx. Online: 24.01.2012.
1
http://www.xsens.com/
BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing
162