Content uploaded by Gustaf Hendeby
Author content
All content in this area was uploaded by Gustaf Hendeby on Aug 31, 2015
Content may be subject to copyright.
Knowledge Exploitation for Human Micro-
Doppler Classification
Cesur Karabacak, Sevgi Z. Gurbuz, Ali C. Gurbuz, Mehmet B. Guldogan, Gustaf Hendeby
and Fredrik Gustafsson
Linköping University Post Print
N.B.: When citing this work, cite the original article.
Cesur Karabacak, Sevgi Z. Gurbuz, Ali C. Gurbuz, Mehmet B. Guldogan, Gustaf Hendeby and
Fredrik Gustafsson, Knowledge Exploitation for Human Micro-Doppler Classification, 2015,
IEEE Geoscience and Remote Sensing Letters, (12), 10.
http://dx.doi.org/10.1109/LGRS.2015.2452311
©2015 IEEE. Personal use of this material is permitted. However, permission to
reprint/republish this material for advertising or promotional purposes or for creating new
collective works for resale or redistribution to servers or lists, or to reuse any copyrighted
component of this work in other works must be obtained from the IEEE.
http://ieeexplore.ieee.org/
Postprint available at: Linköping University Electronic Press
http://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-120433
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, REVISED MAY 2015 1
Knowledge Exploitation for Human Micro-Doppler
Classification
Cesur Karabacak, Sevgi Z. Gurbuz, Member, IEEE, Ali C. Gurbuz, Member, IEEE, Mehmet B.
Guldogan, Member, IEEE, Gustaf Hendeby, Member, IEEE, and Fredrik Gustafsson, Fellow, IEEE
Abstract—Micro-Doppler radar signatures have a great po-
tential for classifying pedestrians and animals, as well as their
motion pattern, in a variety of surveillance applications. Due
to the many degrees of freedom involved, real data needs to be
complemented with accurate simulated radar data to successfully
be able to design and test radar signal processing algorithms.
In many cases, the ability to collect real data is limited by
monetary and practical considerations, whereas in a simulated
environment any desired scenario may be generated. Motion
capture has been used in several works to simulate the human
micro-Doppler signature measured by radar; however, validation
of the approach has only been done based on visual comparisons
of micro-Doppler signatures. This work validates and, more
importantly, extends the exploitation of motion capture data not
just to simulate micro-Doppler signatures but also to use the
simulated signatures as a source of a priori knowledge to improve
the classification performance of real radar data, especially in the
case when the total amount of data is small.
Index Terms—Motion Capture, Human Micro-Doppler, Clas-
sification, Knowledge-Based Signal Processing.
I. INTRODUCTION
THE detection, recognition and classification of human
targets and their activities is a topic of great interest
that has many critical applications, such as search and rescue,
intelligent environments, border control, and security, to name
just a few. Human recognition and classification are typically
accomplished based on discerning differences in the radar
micro-Doppler signature for different targets and activities.
Micro-Doppler refers to the additional frequency modulations
observed about the central Doppler frequency that are caused
not by the gross translational motion of the target, but by vi-
brations or rotations of parts of the target [1]. For example, the
rotation of a helicopter blade, vehicle wheels, and tank tracks,
Manuscript received January 15, 2015.
C. Karabacak is with the Department of Electrical and Electronics Engi-
neering, TOBB University of Economics and Technology, Ankara, Turkey,
as well as the Meteksan Defense Industries, Inc., Ankara, Turkey (e-mail:
ckarabacak@meteksan.com).
S. Z. Gurbuz is with the Department of Electrical and Electronics Engi-
neering, TOBB University of Economics and Technology, Ankara, Turkey, as
well as the TUBITAK Space Technologies Research Institute, Ankara, Turkey
(e-mail: szgurbuz@etu.edu.tr; sevgigurbuz@tubitak.gov.tr).
A. C. Gurbuz is with the Department of Electrical and Electronics En-
gineering, TOBB University of Economics and Technology, Ankara, Turkey
(e-mail: acgurbuz@etu.edu.tr).
M. B. Guldogan is with the Department of Electrical and Electron-
ics Engineering, Turgut Ozal University, Ankara, Turkey (e-mail: bguldo-
gan@turgutozal.edu.tr).
G. Hendeby is with the Department of Electrical Engineering, Linkping
University, Sweden (e-mail: hendeby@isy.liu.se).
F. Gustafsson is with the Department of Electrical Engineering, Linkping
University, Sweden (e-mail: fredrik.gustafsson@liu.se).
as well as the periodic motion of the limbs of animals and
humans all result in visually distinguishable micro-Doppler
signatures. These can then be exploited for automatic target
recognition (ATR) and activity classification.
However, the development and testing of classification algo-
rithms requires data reflecting the wide variety of potential tar-
get characteristics. This includes animals whose signatures are
easily confused with those of humans, such as dogs, donkeys,
cows, and sheep, as well as humans of varying size, build,
and gait that are engaging in different activities at varying
angles, ranges, and velocities. Experimentally generating a
database of this size is both costly and impractical, therefore,
much research, as well as development of commercial systems
involving classification of human micro-Doppler, has utilized
methods for simulating micro-Doppler signatures.
There are two main approaches to simulating human micro-
Doppler signatures: kinematic modeling and motion capture.
The radar return from a human target is generally possible
to represent as the superposition of the returns from point
targets located at varying locations on the human body [2],
[3]. Simulating these returns requires knowledge of the time-
varying positions of each of these point targets, which can
be derived either from kinematic models or from motion
capture data. Some of the kinematic models that have been
utilized in micro-Doppler research studies are quite simplistic,
consisting of just a sinusoidal model of torso motion [4], or
animation of just three parts, the torso and the legs, with a
basic sinusoidal function [5]. The most comprehensive and
ubiquitous kinematic model is the Boulic walking model [6],
which uses a combination of equations and charts to describe
the time-varying motion of 17 points on the human body. It has
been successfully used in a number of human micro-Doppler
studies, e.g. [7]–[9].
The primary disadvantage of kinematic modeling, however,
is that there are no models covering the entire range of
human or animal activity, thus, only special cases can be
simulated. Motion capture data, on the other hand, is derived
from observations of alternative sensors, such as video or
infrared sensors. Typically, a finite number of points on the
human body are marked and their time-varying positions to
the sensor are computed using skeleton-tracking software.
These measurements are then used to simulate the distance
measurements that would have been made had the subject been
observed by radar. Motion capture (MOCAP) based data thus
enables the simulation of any desired sequence of motion for
any subject, including animals, while also incorporating the
individual variations of a subject. Since first being proposed
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, REVISED MAY 2015 2
by Ram, et al. [10] in 2008, motion capture data has been
used for a variety of radar micro-Doppler studies, includ-
ing analysis of the effect of walls [11], [12], the ground
[13], [14], polarization, and operating frequency [14] on the
human micro-Doppler signature. Motion capture techniques
have also been successfully applied to generate radar micro-
Doppler signatures for animals [15]. Indeed, several works
have subsequently utilized motion capture-based simulated
micro-Doppler as a basis for studying the extraction of features
relevant to gait analysis [16] and classification [17]–[20].
However, validation of the approach has primarily been done
by simple visual comparison with measured micro-Doppler
signatures [21]–[24].
In fact, most classification algorithms depend upon accurate
extraction of a set of features from the available data to achieve
high performance. For features derived from simulated data to
yield representative classification results on measured data, the
statistical properties of the features extracted from simulated
data should be comparable to those extracted from real data.
Otherwise, the classification results obtained from simulated
data could be misleading and unrepresentative of the actual
classifier performance.
Beyond validation of simulated data, however, the main goal
of this paper is to show that simulations derived from motion
capture data are indeed accurate enough representations to
be exploited as source of a priori knowledge suitable for
improving micro-Doppler classification performance. More
specifically, this work proposes a novel application of sim-
ulated MOCAP data as a source of training data for real radar
systems, especially in the case when the amount of measured
data is small.
With this aim, in Sec. II, first the radar system used
to collect measured micro-Doppler signatures is described.
Then, in Sec. III, different methods to simulate human micro-
Doppler signatures are discussed, while in Sec. IV the features
extracted from MOCAP data are compared with those of
measured data. Finally, in Sec. V, a novel approach to exploit
simulated micro-Doppler signatures as a priori knowledge for
first training classifiers before testing on measured data is
presented along with classification results.
II. EX PERIM EN TAL MEASUREMENTS OF HUMAN
MIC RO-DOP PL ER SIGNATURES
The experimental radar measurements used in this arti-
cle were captured using a SIRS 77 TD radar, developed by
SAAB AB. It is a frequency modulated continuous wave
(FMCW) radar with linear FM sweeps using the carrier
frequency 77 GHz. The received signal is mixed with the
transmitted signal and only the low bandwidth difference
signal needs to be digitized. This simplifies the hardware
design. The pulse compression (converting to ranges) is done
using a fast Fourier transform (FFT). The range resolution is
approximately 1m. The Doppler filtering utilizes the phase
shift between different sweeps, similar to what is done in
pulsed radars. In the used radar modes, the unambiguous radial
velocity was ensured to be enough for a moving human.
The measurements were conducted on moving persons in
Sweden in the summer 2009, on a day with favorable dry
weather. The experiments were performed on fairly flat ground
covered with short grass, at a place where there were no
buildings to disturb the measurements. The radar was pointing
approximately horizontally. Three different adults (one woman
and two men, of average build and weight) performed the
activities moving either radially towards or away from the
radar (aspect angle 0◦/180◦). Five different activities were
performed: running, jogging, walking, creeping and crawling.
After the acquisition the data was pulse compressed, fol-
lowed by an ideal high-pass filter to remove clutter, and
finally converted to a micro-Doppler signature by using a short
time Fourier transform (STFT) with a Hamming window. The
micro-Doppler spectra from the measurements have previously
been studied and analyzed in [25].
III. HUMAN MIC RO-DOP PL ER SIGNATURES
The micro-Doppler signature for any human activity is the
result of a complex combination of the time-varying motion
of each point on the human body. Suppose the entire human
body is divided into Kparts, which are in turn modeled by
point targets. Then the entire human return, sh, can be written
as the sum of returns from each individual point target [26]:
sh=
K
X
i=1
at,irectˆ
t−td,i
τej−2πfc(t−td,i )+πγ (ˆ
t−td,i)2
.(1)
Here, τis the pulse width; cis the speed of light; γis the
chirp slope; and fcis the transmitted center frequency. The
total time elapsed, t, can be written in terms of the pulse
repetition interval (PRI), T; the fast-time ˆ
t, which denotes the
time as measured from the start of each PRI; and transmitted
pulse number (slow-time), n, as t=T(n−1) + ˆ
t. The time
delay td,i is measured in terms of the total elapsed time tand
is defined as the round-trip travel time between the antenna
and the ith body part. Thus, the time delay is related to the
target range, R, as td= 2R/c. In this way, the received return
may be viewed as a two-dimensional signal that is a function
of fast-time and slow-time. The amplitude, at,i, is defined for
each point target from the range equation as
at,i =Gλ√Ptσi
(4π)3
2R2
i√Ls√La
,(2)
where Gis the antenna gain, λis wavelength, Ptis the
transmitted power, σiis the radar cross-section (RCS) for
the ith body part, Lsrepresents system losses, and Lais
the atmospheric loss. Although this expression includes range
or geometry dependent factors, such as the antenna gain and
atmospheric losses, this work assumes that such factors are
constant for each body part. The RCS is modeled according to
the approximate shape of the body parts, i.e., a sphere for the
head, and ellipsoids for the remaining parts. Thus, grouping
all factors except for the range, Ri, and RCS, σi, of the ith
body part into a constant, A, the amplitude may be expressed
as at,i =Aσi/R2
i.
A. Kinematic Model-Based Signature Simulation
In the literature several different kinematic models have
been used to simulate micro-Doppler. The simplest kinematic
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, REVISED MAY 2015 3
(a) (b) (c) (d) (e)
Fig. 1. (a) The SIRS 77 TD radar used to collect micro-Doppler measurements, and sample walking spectrograms as obtained from (b) measured SIRS77 TD
radar data, (c) Boulic kinematic model, (d) Kinect data, and (e) MOCAP database.
model [27] considers just the oscillation of the torso, which is
represented by a sinusoidal function. Slightly expanding upon
the torso model, the lower-body model [28] comprise three
points, the torso and two legs, where each leg is animated
similarly to a swinging pendulum. The spectrograms of both
these models have some similarity to real data, but important
components of the micro-Doppler signature are absent.
The most widely used model is the Boulic walking model,
which is based on the results of a detailed experimental study
of gait analysis, and uses a combination of equations and
charts to animate 17 different points and joints on the human
body. Thus, the Boulic model is not a closed-form model. All
equations and charts, however, basically depend on just two
physical parameters: the speed at which the person is walking,
and the height of the thigh from the ground (HT).
The most significant limitation of the Boulic model is that
it is only valid for walking, and for speeds in which the
v/HT ratio is less than 2.6 Hz. Moreover, because it is based
on a model, individual gait variations cannot be incorporated
into the simulations. Nevertheless, the resulting micro-Doppler
signatures bear a close similarity to measured data in general
shape, as shown in Fig. 1(a)-(c), with the key exception that it
is much more crisp and clean. The strongest return is due
to the torso, which in both signatures appears as a strong
sinusoid around the central Doppler shift that corresponds to
the average radial velocity of the person. Limb motion appears
as larger amplitude oscillations, the largest being that of the
legs.
B. Motion Capture-Based Signature Simulation
More realistic human spectrograms may be obtained by
exploiting video human motion capture data, from which the
time-varying ranges of points on the human body can be
computed. High quality motion capture is typically accom-
plished by placing infrared markers upon the human body and
observing these positions with a camera. Commercial motion
capture sensor suites are available, but these are typically
quite expensive ($30,000–$60,000). A low-cost system using
the Kinect sensor has recently been developed [29], but the
accuracy is less than that of commercial systems.
Databases storing motion capture data are available [30]–
[32] and have been used for micro-Doppler signature simu-
lation. For example, the Carnegie Mellon University (CMU)
Motion Research Laboratory has developed a freely distributed
library of human motion capture data [32]. It includes a
wide range of activities, such as walking, running, crawling,
jumping, and boxing, to name just a few. The data was
collected with the aid of 41 markers placed on the human
body and was recorded by 12 infrared cameras at a frequency
of 120 Hz. The database contains a total of 2,605 different
motion records belonging to 112 different subjects.
The primary disadvantage of using MOCAP databases is
that the test subjects used and test scenarios recorded are not
in the control of the user. This deficiency can be remedied
by using Kinect, however, the signature quality is poorer than
that of the CMU Motion Capture Library, as can be seen from
Fig. 1(d) and (e).
IV. EVALUATIO N OF SI MULATED SIGNATURES
The performance of classification algorithms is highly de-
pendent upon the features extracted from the micro-Doppler
signatures. Thus, in this section, CMU MOCAP spectrograms
are compared with measured SIRS 77 TD radar data based
upon the distributions of features used for classification.
A. Feature Extraction
Feature extraction, one of the most critical steps in a
classification system, is the process of computing numerical
indicators (features) that will enable the discrimination of
different classes of data. Moreover, feature extraction enables
the classification process to progress using a small set of
features rather than make computations on the much larger
set of raw data. In this way, the computation time of the
classification process is also reduced.
Proper feature selection is important to optimize classifica-
tion performance. In this study, three features (the mean of
the torso velocity, the variance of the upper envelope, and
the variance of the lower envelope) [33] are extracted from
micro-Doppler spectrograms. The mean of the torso velocity
is a representative measurement of the translational motion of
the target. The variances of the upper and the lower envelopes
reflect movements of arms and legs of the subject. These
three features are together an effective set to define different
human activities and can be computed from the spectrogram,
as illustrated in Fig. 2.
The envelopes of the spectrograms are extracted using a
percentile technique proposed by Van Dorp and Groen [34]
in 2008. First, the element of the envelope in each column of
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, REVISED MAY 2015 4
(a) (b)
Fig. 2. Envelope extraction from spectrograms: (a) original spectrogram, and
(b) extracted envelopes.
the spectrogram data is computed. To do this, the cumulative
amplitude distribution for each column is computed:
P(v, t) =
v
X
v=vmin
s(v, t)vmax
X
v=vmin
s(v, t)(3)
where vis the radial velocity, tis time, and sthe matrix storing
the values for amplitude (power) of the spectrogram. For each
column, the value P(U(t), t) = 0.97 is used to calculate the
upper envelope; P(C(t), t) = 0.5is used to calculate the
central envelope; and P(L(t), t)=0.03 is used to calculate
the lower envelope. Once these three envelopes are extracted,
their average value is computed, representing three different
features in a three dimensional feature space.
B. Statistical Comparison of the Features
The three features introduced in the previous subsection
are extracted from both measured and MOCAP data. In
both databases, there are four different human motion types:
walking, running, crawling, and creeping. The duration of each
data file is 1.5s and the subjects move towards radar during
the data acquisition process. Distribution of data types in these
two databases is given in Table I.
To compare the distribution of features obtained from MO-
CAP data with that of real data, scatter graphs of features
from each data set are presented in Fig. 3 and 4, respectively.
In both data sets, running is clearly separable from the other
classes of motion, while features appear to have similar
values for creeping and crawling, limiting their discriminating
power. Walking appears to be more easily distinguishable
from MOCAP-based simulated spectrograms than would be
possible from real measured data.
Classification performance attained using a nearest neighbor
(NN) classifier also validates the comparability of MOCAP-
derived signatures with measured signatures. Using 84 MO-
CAP signatures for training, 25 of the remaining 28 MOCAP
signatures were correctly classified, with just three crawling
signatures confused with creeping. In comparison, 15 of 16
measured SIRS 77 TD radar signatures were correctly classi-
fied when 40 measured signatures were used as training.
V. SI MU LATED SI GNATUR ES AS PRIO R KNOWL ED GE
To date, simulated micro-Doppler signatures have primarily
been used just as a means to approximate real data in per-
formance testing. In fact, simulated signatures represent prior
knowledge that can be exploited to improve the classification
TABLE I
DISTRIBUTION OF DATA TYPES
CLASSES
Walking Running Crawing Creeping
MOCAP 28 28 28 28
Measured 18 15 11 12
Fig. 3. Scatter graph of the features for the MOCAP database.
Fig. 4. Scatter graph of the features for the SIRS 77 TD radar.
performance of measured data. An important factor affecting
performance is the amount of data used to train the classifier.
Increasing the size of the training set tends to increase the
rate of correct classification. However, in real operational
scenarios, the total data set size may not be that large, limiting
the ability to train the classifier. In this work, this constraint
is overcome by using simulated data to train measured data.
To assess the potential performance attainable by training
with MOCAP data, classification of the SIRS 77 TD radar data
is accomplished in two ways: 1) a subset of measured data
is used to train a classifier to test the remaining measured
data; and 2) MOCAP-derived signatures are used to train all
measured data. With the first approach, it was found that
a correct classification rate of 94% was achieved when at
least 70% of the data was used for training. When MOCAP
data was used to train the classifier, the highest classification
performance attained was 93% when all 112 MOCAP-derived
signatures were used for training (Fig. 5). Confusion matrices
corresponding to the best classification results are shown in
Table II, as well as classification metrics in Table III.
VI. CONCLUSION
This work validates the use of MOCAP-derived simulated
data through the examination of the statistical distributions
of features, as well as through comparison of classification
performance. Moreover, a novel application of using simulated
MOCAP data to train classifiers for testing real data is
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, REVISED MAY 2015 5
Fig. 5. Dependency of classification performance on number of MOCAP-
signatures used for training. Vertical bars denote confidence interval (standard
deviation) of results.
TABLE II
CLASSIFICATION PERFORMANCE OF N N ON MOCAP-BAS ED A ND
SIRS 77 TD MEASURED RA DAR DATA
Walking Running Crawing Creeping
Training with SIRS77 TD Measured Radar Data
Walking 5 0 0 0
Running 0 4 0 0
Crawling 0 0 3 0
Creeping 0 0 1 3
Training with MOCAP-Based Simulated Data
Walking 18 0 0 0
Running 0 15 0 0
Crawling 0 0 11 0
Creeping 0 0 4 8
TABLE III
COMPARISON OF CLASSIFICATION METRICS
Walking Running Crawing Creeping
Training with SIRS77 TD Measured Radar Data
Sensitivity 1 1 1 0.75
Specificity 1 1 0.92 1
Precision 1 1 0.75 1
Training with MOCAP-Based Simulated Data
Sensitivity 1 1 1 0.67
Specificity 1 1 0.91 1
Precision 1 1 0.73 1
proposed. This method can be applied using different MOCAP
data sets, as well as different features and classifiers. Generally
including as many different individuals as possible in the
training set mitigates any potential losses due to different styles
of gait. In summary, results show that simulated data can yield
comparable results to that as achieved with training on real
data. This result is especially significant for the classification
of small data sets, when the amount of training data is
insufficient.
ACKNOWLEDGMENT
The authors gratefully acknowledge support from SAAB
and funding from Security Link, the EU FP7 Project No.
PIRG-GA-2010-268276 and TUBITAK Career No. 113E105.
REFERENCES
[1] V. Chen, The micro-Doppler effect in radar, Artech House, 2011.
[2] J.L. Geishemer, W.S. Marshall, E. Greneker, A continuous-wave (CW)
radar for gait analysis, in Proc. 25th Asilomar Conf. on Signals, Systems
and Computers, pp. 834-838, 2001.
[3] P. Van Dorp, and F.C.A. Groen, Human walking estimation with radar,
IEE Radar Sonar Navig.,Vol. 150, No. 5, pp 356365, Oct. 2003.
[4] G.E. Smith, K. Woodbridge, and C.J. Baker, Multistatic micro-doppler
signature of personnel, in Proc. IEEE Radar Conference, 2008.
[5] D. Tahmoush, and J. Silvious, Simplified model of dismount micro-
Doppler and RCS, in Proc. IEEE Radar Conference, 2010.
[6] R. Boulic, N.M. Thalmann, and D. Thalmann, D. A global human walking
model with real-time kinematic personification, Vis. Comput., Vol. 6, No.
6, pp 344358, Nov. 1990.
[7] T. Damarla, M. Bradley, A. Mehmood, J.M. Sabatier, J.M., Classification
of Animals and People Ultrasonic Signatures, IEEE Sensors Journal.
[8] S.R. Groot, A.G. Yarovoy, R.I.A. Harmanny, J.N. Driessen, Model-based
classification of human motion: Particle filtering applied to the Micro-
Doppler spectrum, in Proc. EuRAD, 2012.
[9] I. Bilik, and J. Tabrikian, ”Radar target classification using Doppler
signatures of human locomotion models,” IEEE Trans. AES, Vol.43, No.4,
pp.1510-1522, October 2007.
[10] S.S. Ram, H. Ling, ”Simulation of human microDopplers using com-
puter animation data,” IEEE Radar Conference, 2008.
[11] S.S. Ram, C. Christianson, H. Ling, ”Simulation of high range-resolution
profiles of humans behind walls,” IEEERadar Conf., 2009.
[12] S.S. Ram, C. Christianson, Y. Kim, H. Ling, ”Simulation and Analysis
of Human Micro-Dopplers in Through-Wall Environments,” IEEE Trans.
Geoscience and Remote Sensing, vol.48, no.4, pp.2015-2023, April 2010.
[13] S.S. Ram, R. Bhalla, H. Ling, ”Simulation of human radar signatures
in the presence of ground,” IEEE Antennas and Propagation Society
International Symposium, 1-5 June 2009.
[14] J. Park, J.T. Johnson, N. Majurec, M. Frankford, E. Culpepper, J.
Reynolds, J. Tenbarge, L. Westbrook, ”Software defined radar studies
of human motion signatures,” IEEE Radar Conference, 2012.
[15] S.S. Ram, and H. Ling, ”Microdoppler signature simulation of computer
animated human and animal motions,” IEEE Antennas and Propagation
Society International Symposium, 2008.
[16] A. Ghaleb, L. Vignaud, J-M. Nicolas, ”Micro-Doppler analysis of
pedestrians in ISAR imaging,” IEEE Radar Conference, 2008.
[17] L. Fei, H. Binke, Z. Hang, D. Hao, ”Human gait recognition using micro-
Doppler features,” 12th Global Symp. on Millimeter Waves (GSMM),
pp.326-329, 27-30 May 2012.
[18] S.Z. Gurbuz, B. Tekeli, M. Yuksel, et. al., ”Importance ranking of
features for human micro-Doppler classification with a radar network,”
IEEE Int. Conf. on Information Fusion, 9-12 July 2013.
[19] R.G. Raj, V.C. Chen, R. Lipps, ”Analysis of radar human gait signa-
tures,” Signal Processing, IET , vol.4, no.3, pp.234,244, June 2010.
[20] Y. Yang, W. Zhang, and C. Lu, ”Classify human motions using micro-
Doppler radar.” SPIE Defense and Security Symp., 2008.
[21] V.C. Chen, W.J. Miceli, B. Himed, ”Micro-Doppler analysis in ISAR -
review and perspectives,” Int.Radar Conference, 12-16 Oct. 2009.
[22] V.C. Chen, ”Doppler signatures of radar backscattering from objects
with micro-motions,” IET Signal Processing, 2, (3), pp.291-300, 2008.
[23] C. Karabacak, S.Z. Gurbuz, A.C. Gurbuz, ”Radar simulation of human
micro-Doppler signature from video motion capture data,” IEEE Signal
Processing and Communications Applications Conference, 2013.
[24] L. Vignaud, A. Ghaleb, J. Le Kernec, J-M. Nicolas, ”Radar high
resolution range and micro-Doppler analysis of human motions,” IEEE
Int.Radar Conference, 12-16 Oct. 2009.
[25] S. Bjorklund, H. Petersson, A. Nezirovic, M.B. Guldogan, F. Gustafsson,
Millimeter-Wave Radar Micro-Doppler Signatures of Human Motion,
International Radar Symposium, Sept. 2011.
[26] W.G. Carrara, R.S. Goodman, R.M. Majewski, Spotlight Synthetic
Aperture Radar, Artech House, Boston, 1995.
[27] D. Tahmoush, and J. Silvious, Simplified model of dismount mi-
croDoppler and RCS, in Proc. IEEE Radar Conference, 2010.
[28] G.E. Smith, K. Woodbridge, and C.J. Baker, Multistatic micro-Doppler
signature of personnel, in Proc. IEEE Radar Conference, 2008.
[29] B. Erol, C. Karabacak, S.Z. Gurbuz, A Kinect-based human micro-
Doppler simulator, Aerospace and Electronic Systems Mag., accepted.
[30] SNU Movement Research Lab [Online]. Available: http://mrl.snu.ac.kr/
[31] ACCAD Motion Capture Lab. [Online]. Available: http://accad.osu.edu/
research/mocap/mocap data.htm
[32] CMU Graphics Lab Motion Capture Database [Online]. Available: http:
//mocap.cs.cmu.edu/
[33] Y. Kim and H. Ling, ”Human activity classification based on micro-
Doppler signatures using a support vector machine,” IEEE Trans. Geo-
science and Remote Sensing, Vol. 47, Iss. 5, pp. 1328 - 1337, May 2009.
[34] P. Van Dorp, F.C.A. Groen, ”Feature-based human motion parameter
estimation with radar,” IET Radar, Sonar and Navigation, Vol.2, No.2,
pp.135-145, April 2008.