PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

We introduce a simple but effective technique in automatic hand gesture recognition using radar. The proposed technique classifies hand gestures based on the envelopes of their micro-Doppler signatures. These envelopes capture the distinctions among different hand movements and their corresponding positive and negative Doppler frequencies which are generated during each gesture act. We detect the positive and negative envelopes separately, and form a feature vector of their augmentation. We use the $k$-nearest neighbor ($k$NN) classifier and Manhattan distance (L1) measure, in lieu of Euclidean distance (L2), so as not to diminish small but critical envelope values. It is shown that this method outperforms both low-dimension representation techniques based on principal component analysis (PCA) and sparse reconstruction using Gaussian-windowed Fourier dictionary, and can achieve very high classification rates.
Content may be subject to copyright.
Hand Gesture Recognition based on Radar
Micro-Doppler Signature Envelopes
Moeness G. Amin
Center for Advanced Communications
Villanova University
Villanova, PA 19085, USA
moeness.amin@villanova.edu
Zhengxin Zeng, Tao Shan
School of Information and Electronics
Beijing Institute of Technology
Beijing, China
{3120140293,shantao}@bit.edu.cn
Abstract—We introduce a simple but effective technique in
automatic hand gesture recognition using radar. The proposed
technique classifies hand gestures based on the envelopes of
their micro-Doppler signatures. These envelopes capture the
distinctions among different hand movements and their corre-
sponding positive and negative Doppler frequencies which are
generated during each gesture act. We detect the positive and
negative envelopes separately, and form a feature vector of their
augmentation. We use the k-nearest neighbor (kNN) classifier and
Manhattan distance (L1) measure, in lieu of Euclidean distance
(L2), so as not to diminish small but critical envelope values.
It is shown that this method outperforms both low-dimension
representation techniques based on principal component anal-
ysis (PCA) and sparse reconstruction using Gaussian-windowed
Fourier dictionary, and can achieve very high classification rates.
Keywords—Hand gesture recognition; time-frequency repre-
sentations; micro-Doppler signature envelope.
I. INTRODUCTION
Radar systems assume an important role in several areas of
our daily life, such as air traffic control, speed enforcement
systems, and advanced driver assistance systems [1–3]. Re-
cently, radar has also become of increased interest for indoor
applications. In particular, human activity monitoring radar
systems are rapidly evolving with application that include gait
recognition, fall motion detection for elderly care and aging-
in-place technologies [4, 5].
Over the past decade, much work has been done in hu-
man motion classifications which include daily activities of
walking, kneeling, sitting, standing, bending, falling, etc. [6–
16]. Distinguishing among the different motions is viewed
as an inter-class classification [6–13], whereas the intra-class
classification amounts to identifying the different members of
the same class, e.g., classifying normal and abnormal gaits
[14–16]. There are two main approaches of human motion
classifications, namely those relying on handcrafted features
that relate to human motion kinematics [7–9, 14], and others
which are data driven and include low-dimension represen-
tations [6, 15], frequency-warped cepstral analysis [13], and
neural networks [10–12, 16].
The work of Mr. Zhengxin is funded by the International Graduate
Exchange Program of Beijing Institute of Technology, and was performed
while he was a Visiting Scholar at the Center for Advanced Communications,
Villanova University
In addition to classifying human motions, radars have been
recently used for gesture recognition which is an important
problem in a variety of applications that involve smart homes
and human-machine interface for intelligent devices [17–21].
The latter is considered vital in aiding the physically impaired
who might be wheelchair confined or bed-ridden patients.
The goal is to enable these individuals to be self-supported
and independently functioning. In essence, automatic hand
gesture recognition is poised to make our homes more user
friendly and most efficient through the use of contactless
radio frequency (RF) sensors that can identify different hand
gestures for instrument and household appliance control.
The same approaches employed for classifying human daily
activities can be applied for recognition of hand gestures using
the electromagnetic (EM) sensing modality. However, there is
an apparent difference between micro-Doppler signatures of
hand gestures and those associated with motion activities that
involve human body. Depending on the experiment and data
collection radar specs and setting, Micro-Doppler representa-
tions of hand gestures can be simple, limited to short time
duration and small bandwidth, and are mainly characterized
by their confined power concentrations in the time-frequency
domain. On the other hand, the micro-Doppler signatures of
body motions are intricate, of multi-component signals, span
relatively longer time periods and assume higher Doppler
frequencies.
In this paper, we present a method to discriminate five
classes of dynamic hand gestures using radar micro-Doppler
sensor. These classes are swiping hand, hand rotation, flip-
ping fingers, calling and snapping fingers. We begin with
many types of hand motions, and use the canonical angle
metric to assess the subspace similarities constructed from
their respective time-frequency distributions [22, 23]. Based
on the results, we group these motions into the five most
dissimilar classes. Two micro-Doppler features are extracted
from the data spectrograms. They correspond to the upper and
lower envelopes of the hand gesture micro-Doppler signatures.
The two envelopes implicitly capture, for each motion, the
positive-negative frequency differences, the time alignments
and misalignments of the peak positive and negative Doppler
frequencies, and the signature extent and occupancy over the
joint time and frequency variables.
We compare the proposed approach with that based on
PCA [6, 15] and on sparse reconstruction employing Gaussian-
windowed Fourier dictionary [17]. In the latter, the classifier
was applied to hand gesture data showing signatures com-
prising rather detached power concentrated time-frequency
regions The experimental results applied to our measured data
demonstrate that the proposed method outperform the above
two methods, and achieve a classification accuracy higher than
96%.
The remainder of this paper is organized as follows. In Sec-
tion II, we present the extraction method of micro-Doppler sig-
nature envelopes and discusses the employed classifier. Section
III describes the radar data collection and pre-processing of
hand gestures. Section IV gives the experimental results based
on the real data measurements. Section V is the conclusion of
the paper.
II. HA ND GE ST UR E RECOGNITION ALGORITHM
A. Time-frequency Representations
Hand gestures generate non-stationary radar back-scattering
signals. Time-frequency representations (TFRs) are typically
employed to analyze these signals in the joint-variable do-
mains, revealing what is referred to as micro-Doppler signa-
tures. A typical technique of TFRs is the spectrogram. For a
discrete-time signal s(n)of length N, the spectrogram can be
obtained by taking the short-time Fourier transform (STFT)
S(n, k) =
L1
X
m=0
s(n+m)h(m)ej2πmk
N
2
(1)
where n= 0,· · · , N 1is the time index, k= 0,· · · K1
is the discrete frequency index, and Lis the length of the
window function h(·).
B. Extraction of the Micro-Doppler Signature Envelopes
We select features specific to the nominal hand gesture
local frequency behavior and power concentrations. These
features are the upper and lower envelope in the spectrograms.
The envelopes attempt to capture, among other things, the
maximum positive frequency and negative frequencies, length
of the event and its bandwidth, the relative emphases of the
motion towards and away from the radar, i.e., positive and
negative Doppler frequencies. In essence, the envelopes of
the signal power concentration in the time-frequency domain
may uniquely characterize the different hand motions. The
envelopes of the micro-Doppler signature can be determined
by an energy-based thresholding algorithm [24]. First, the
effective bandwidth of each gesture frequency spectrum is
computed. This defines the maximum positive and negative
Doppler frequencies. Second, the spectrogram S(n, k)is di-
vided into two parts, the positive frequency part and the
negative frequency part. The corresponding energies of the two
parts, EU(n)and EL(n), at slow-time are computed separately
as
EU(n) =
K
21
X
k=0
S(n, k)2(2)
EL(n) =
K1
X
k=K
2
S(n, k)2(3)
These energies are then scaled to define the respective thresh-
olds, TUand TL,
TU(n) = EU(n)·σU(4)
TL(n) = EL(n)·σL(5)
where σUand σLrepresent the scale factors, both are less
than 1. These scalars can be chosen empirically. However, an
effective way for their selections is to maintain the ratio of
the energy value to the threshold value constant over all time
samples. For the upper envelope, this ratio can be computed
by finding both values at the maximum positive Doppler
frequency. Once the threshold is computed per equation (4),
the upper envelope is then found by locating the Doppler
frequency for which the spectrogram assumes equal or higher
value. Similar procedure can be followed for the lower enve-
lope. The upper envelope eU(n)and lower envelope eL(n),
are concatenated to form a long feature vector e= [eU, eL].
C. Classifier
We apply proper classifiers based on the envelope features
extracted from the spectrograms. The kNN and Support vector
Machine (SVM) are among the most commonly used classi-
fiers in pattern recognition which are used in this paper. In
particular, the kNN is a simple machine learning classification
algorithm. For each test sample, the algorithm calculates the
distance to all the training samples, and selects the kclosest
training samples. Classification is performed by assigning the
label that is most frequent among these samples [25]. Clearly,
the best choice of kwould depend on the data. In this work,
kis set to 1. Four different distance metrics are considered,
namely, the Euclidean distance, the Manhattan distance [26],
the Earth Mover’s distance (EMD) [27] and the modified
Hausdorff distance (MHD) [28].
SVM is a supervised learning algorithm [29]. It exhibits
clear advantages in nonlinear and high dimension problems.
III. HAND GES TU RE SUBCLASSES
The data analyzed in this paper was collected in the Radar
Imaging Lab at the Center for Advanced Communications,
Villanova University. The radar system used in the experi-
ment generates continuous wave, with carrier frequency and
sampling rate equal to 25 GHz and 12.8 kHz, respectively.
The radar was placed at the edge of a table. The hand was
approximately 20 cm away from radar at zero angle, and the
arm remained fixed as much as possible during each gesture
motion.
As depicted in Fig.1. The following 15 hand gestures were
conducted: (a) Swiping hand from left to right, (b) Swiping
hand from right to left, (c) Swiping hand from up to down,
(d) Swiping hand from down to up, (e) Horizontal rotating
hand clockwise, (f) Horizontal rotating hand counterclockwise,
(g) Vertical rotating hand clockwise, (h) Vertical rotating
hand counterclockwise, (i) Opening hand, (j) Flipping fingers,
(k) Clenching hand, (l) Calling, (m) Swipe left with two
fingers, (n) Snapping fingers, (o) Pinching index. Four persons
participated in the experiment. Each hand gesture was recorded
over 8 seconds to generate one data segment. The recording
was repeated for 5 times. Each data segment contained 2 or
3 individual hand motions, and a 1 second time window is
applied to capture the individual motions. As such, repetitive
motions and associated duty cycles were not considered in
classifications. In total, 755 segments of data for 15 hand
gestures were generated.
(a) (b) (c) (d) (e)
(f) (g) (h) (i) (j)
(k) (l) (m) (n) (o)
Fig. 1. Illustrations of 15 different hand gestures.
Fig. 2 shows examples of spectrograms and corresponding
envelopes for different hand gestures. The employed sliding
window h(·)is rectangular with length L=2048 (0.16 s),
and Kis set to 4096. It is clear that the envelopes can well
capture the salient features of the respective spectrograms. It
is also evident that the micro-Doppler characteristics of the
spectrograms are in agreement and consistent with each hand
motion kinematics. For example, for the hand gesture ‘Swiping
hand’, the hand moves closer to the radar at the beginning
which causes the positive frequency, and then moving away
from the radar which induces the negative frequency.
Observing the spectrograms in Fig. 2, it is noticeable that
similar motions generate similar signatures. To mathematically
confirm these resemblances, we consider sub-grouping the 15
hand gestures using the Canonical correlation measure [22]. In
this case, the spectrograms are converted to gray-scale images
with the size 100 ×100, and then vectorized with the size
1×10000.
Define matrix Xcontains Mvectorized images Si, i =
1,· · · , M of a specific hand gesture,
X= [x1|x2|···|xM](6)
The d-dimensional subspace of a specific hand gesture can
be obtained by taking PCA of X[30]. Suppose Φ1and Φ2are
two d-dimensional linear subspaces, the canonical correlations
of the two subspaces are the cosines of principal angles, and
are defined as [31]:
cos θi= max
uiΦ1
max
viΦ2
uT
ivi(7)
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j)
(k) (l) (m)
(n) (o)
Fig. 2. Spectrograms and corresponding envelopes of 15 different hand
gestures.
subject to ||u|| =||v|| = 1, uiTuj=viTvj= 0, i 6=j. Let U
and Vdenote unitary orthogonal bases for two subspaces, Φ1
and Φ2. The singular value decomposition (SVD) of UTVis
UTV=PΛQ(8)
The canonical correlations are the singular values Λ, i.e.,
cos(θi) = λi, i = 1,· · · , d. The minimum angle is used
to measure the closeness of two subspaces. Table I shows
the canonical correlations coefficients, from which we can
clearly see the similarities between the different hand gestures;
larger coefficient indicates the two hand gestures are more
similar. The red part of the table indicates the coefficient
exceeds 0.9, and the yellow part means the coefficient is over
0.85. According to the Table numbers, we group the 15 hand
gestures into 5 Class. Class I is the gesture ‘Swiping hand’
which contains motions (a), (b), (c) and (d). Class II represents
the gestures ‘Hand rotation’ which contains motions (e), (f),
(g) and (h). The gesture ‘Flipping fingers,’ which involves
motions (i) and (j), makes Class III. Class IV is the gesture
‘Calling,’ which has motions (k), (l) and (m). The last Class
V is the gesture ‘Snapping fingers’; it has motions (n) and
(o). It is important to note that other similarity measures [*]
can be applied, in lieu of the canonical correlation. However,
we found the canonical correlation most consistent with the
visual similarities.
IV. EXPERIMENTAL RESU LTS
In this section, all 755 data segments are used to validate the
proposed method where 70% of the data are used for training
and 30% for testing. The classification results are obtained
TABLE I. CANONICAL CORRELATIONS COEFFI CIE NT S
b c d e f g h i j k l m n o
a0.79 0.83 0.91 0.70 0.75 0.79 0.84 0.69 0.66 0.78 0.77 0.76 0.77 0.81
b0 0.92 0.80 0.70 0.68 0.82 0.82 0.65 0.61 0.78 0.82 0.83 0.73 0.60
c0 0 0.76 0.64 0.59 0.85 0.88 0.72 0.65 0.80 0.80 0.82 0.76 0.69
d0 0 0 0.61 0.68 0.81 0.75 0.57 0.55 0.78 0.67 0.60 0.63 0.64
e0 0 0 0 0.86 0.70 0.75 0.59 0.66 0.56 0.72 0.66 0.72 0.71
f0 0 0 0 0 0.78 0.83 0.70 0.70 0.67 0.73 0.70 0.78 0.79
g0 0 0 0 0 0 0.85 0.67 0.67 0.78 0.66 0.71 0.74 0.73
h0 0 0 0 0 0 0 0.55 0.60 0.72 0.67 0.61 0.71 0.71
i0 0 0 0 0 0 0 0 0.87 0.75 0.61 0.67 0.76 0.74
j0 0 0 0 0 0 0 0 0 0.68 0.61 0.68 0.83 0.73
k0 0 0 0 0 0 0 0 0 0 0.94 0.94 0.83 0.76
l0 0 0 0 0 0 0 0 0 0 0 0.93 0.73 0.66
m0 0 0 0 0 0 0 0 0 0 0 0 0.77 0.63
n0 0 0 0 0 0 0 0 0 0 0 0 0 0.82
by 1000 Monte Carlo trials. Three different automatic hand
gesture approaches are compared with the proposed method.
These are: 1) the empirical feature extraction method; 2) the
PCA-based method [15]; 3) the sparse reconstruction-based
method [17].
A. Empirical Feature Extraction Method
Three empirical features are extracted from the spectro-
grams to describe the hand gestures motions, namely the
length of the event, the ratio of positive-negative frequency
and the signal bandwidth. Fig. 3 is an example showing these
handcrafted features.
Fig. 3. Empirical feature extraction.
1) Length of the event T:This describes the effective time
duration to perform each hand gesture,
T=tets(9)
where tsand terepresent the start time and the end time of a
single hand gesture, respectively.
2) Ratio of positive-to-negative peak frequencies R:This
feature is obtained by finding ratio of the maximum positive
frequency value,fp, and maximum negative frequency value,
fn,
R=
fp
fn
(10)
where |·| is the absolute function.
3) Bandwidth Bw:This is a measure of the the signal
effective width,
Bw=|fp|+|fn|(11)
The scatter plot of the above extracted features is shown in
Fig. 4. Table II depicts the nominal behavior of these values
over the different classes considered. When using kNN-L1 as
the classifier, the recognition accuracy based on these features
is only 68% with the confusion matrix shown in Table II.
Fig. 4. Scatter plot of three extracted empirical features.
TABLE II. CONFUSION MATRIX YIE LDE D BY EMPIRICAL FE ATURE EX-
TR ACTI ON ME THO D
I II III IV V
I66.79% 13.80% 4.08% 9.18% 6.15%
II 20.04% 64.65% 3.53% 4.88% 6.90%
III 9.94% 5.59% 76.53% 0.03% 7.91%
IV 19.04% 6.74% 0.65% 71.79% 1.78%
V12.03% 10.96% 12.28% 11.59% 53.14%
B. Proposed Envelope-based Method
As discussed in Section II, the extracted envelopes are fed
into the kNN classifier, with different distance measures, and
SVM classifier. The recognition accuracy are presented in
Table III. It is cleat that the kNN classifier based on L1
distance achieves the highest accuracy, over 96%, followed
by those employing the modified Hausdorff distance and
the Euclidean distance. Different from other distances, the
L1 distance attempts to properly account for small envelope
values. The confusion matrix of the kNN classifier based on
the L1 distance is shown in Table IV, from which we can find
that Class III and Class IV are most distinguishable with the
accuracy over 98%.
TABLE III. RECOGNITION AC CUR ACY W ITH DI FFE REN T TYP ES OF CL AS -
SI FIER
Accuracy
SVM 83.07%
kNN-L1 95.23%
kNN-L2 93.87%
kNN-EMD 81.51%
kNN-MHD 93.95%
TABLE IV. CONFUSION MATRIX YIE LD ED BY ENVELOPE ME THO D
BAS ED O N kNN-L1 CLAS SI FIER
I II III IV V
I95.23% 3.17% 0.14% 1.46% 0
II 3.03% 95.39% 0.01% 0.06% 1.51%
III 0.07% 0 99.01% 0.28% 0.64%
IV 0.61% 0 1.16% 98.21% 0.02%
V0 2.31% 2.61% 2.83% 92.25%
C. PCA-based Method
For the PCA-based method, each sample represents a spec-
trogram image of 100 ×100 pixels. The number of principal
TABLE V. C ONFUSION MATRI X YIEL DE D BY PCA-BA SED ME THO D WI TH
d= 30
I II III IV V
I89.50% 3.02% 0.67% 6.80% 0.01%
II 2.92% 94.83% 0 1.45% 0.80%
III 2.85% 1.23% 94.42% 0 1.50%
IV 5.24% 0.25% 1.37% 93.14% 0
V3.24% 8.14% 5.03% 1.83% 81.76%
components dis determined by the eigenvalues. Fig.5 shows
how the classification accuracy changes with d, with the
recognition rate increases as dincreases. However, there is
no significant improvement of the recognition accuracy past
d= 30. Table V is the confusion matrix using 30 eigenvalues.
Although the PCA method can achieve an overall accuracy of
92.71%, it is clearly outperformed by the proposed method.
Fig. 5. Performance of PCA with different number of principal components.
D. Sparsity-based Method
The features used for this method are the time-frequency
trajectories. Details of the sparsity-based method can be found
in [17]. The trajectory consists of three parameters, namely
the time-frequency position (ti, fi), i = 1,· · · , whereP and
the intensity Ai,Pis the sparsity level that is set to 10
in this paper. Hence, each sample contains 30 features. The
spectrograms of reconstructed signals and the Plocations
of time-frequency trajectory are plotted in Fig. 6 and Fig.
7. In the training process, the K-means algorithm is used
to cluster a central time-frequency trajectory [32]. In the
testing process, the kNN classifier based on the modified
Hausdorff distance is applied to measure the distance between
the testing samples and central time-frequency trajectories.
The corresponding confusion matrix is presented in Table VI.
The overall recognition accuracy was found to be only about
70% when applied to our data.
TABLE VI. C ONFUSION MATRIX YIELD ED BY SPARSITY-BASED ME TH OD
I II III IV V
I71.72% 11.36% 1.45% 11.74% 3.73%
II 10.95% 81.29% 2.57% 0.28% 4.91%
III 7.40% 2.10% 83.63% 0.69% 6.18%
IV 16.04% 6.52% 1.22% 74.14% 2.08%
V6.65% 15.05% 9.96% 10.02% 58.32%
V. CONCLUSION
We introduced a simple but effective technique for auto-
matic hand gesture recognition based on radar micro-Doppler
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j)
(k) (l) (m)
(n) (o)
Fig. 6. Spectrograms of reconstructed signals with P= 10.
(a) (b) (c) (d)
(e) (f) (g) (h)
(i) (j)
(k) (l) (m)
(n) (o)
Fig. 7. Locations of timefrequency trajectories with P= 10.
signature envelopes. An energy-based thresholding algorithm
was applied to separately extract the upper (positive) envelope
and the lower (negative) envelope of the signal spectrogram.
We used the canonical correlation coefficient to group 15
hand gestures into five different classes. The members of
each class have close signature behavior. The extracted en-
velopes were concatenated and inputted to different types
of classifiers. It was shown that the kNN classifier based
on L1 distance achieves the highest accuracy and provided
over 96 percent classification rate. The experimental results
also demonstrated that the proposed method outperformed
the lower dimensional PCA-based method, the sparsity-based
approach using Gaussian-windowed Fourier dictionary, and
existing techniques based on handcrafted features.
REFERENCES
[1] V. D. Hopkin, Human factors in air traffic control. CRC Press, 2017.
[2] L. Du, Q. Sun, C. Cai, J. Bai, Z. Fan, and Y. Zhang, “A vehicular mobile
standard instrument for field verification of traffic speed meters based
on dual-antenna doppler radar sensor,Sensors, vol. 18, no. 4, p. 1099,
2018.
[3] J. Zolock, C. Senatore, R. Yee, R. Larson, and B. Curry, “The use
of stationary object radar sensor data from advanced driver assistance
systems (adas) in accident reconstruction,” SAE Technical Paper, Tech.
Rep., 2016.
[4] M. Amin, Radar for Indoor Monitoring: Detection, Classification, and
Assessment. CRC Press, 2017.
[5] M. G. Amin, Y. D. Zhang, F. Ahmad, and K. D. Ho, “Radar signal
processing for elderly fall detection: The future for in-home monitoring,”
IEEE Signal Processing Magazine, vol. 33, no. 2, pp. 71–80, 2016.
[6] B. Jokanovic, M. Amin, F. Ahmad, and B. Boashash, “Radar fall detec-
tion using principal component analysis,” in Radar Sensor Technology
XX, vol. 9829. International Society for Optics and Photonics, 2016,
p. 982919.
[7] P. Van Dorp and F. Groen, “Feature-based human motion parameter
estimation with radar,IET Radar, Sonar & Navigation, vol. 2, no. 2,
pp. 135–145, 2008.
[8] Y. Kim, S. Ha, and J. Kwon, “Human detection using doppler radar
based on physical characteristics of targets,” IEEE Geoscience and
Remote Sensing Letters, vol. 12, no. 2, pp. 289–293, 2015.
[9] B. G. Mobasseri and M. G. Amin, “A time-frequency classifier for
human gait recognition,” in Optics and Photonics in Global Homeland
Security V and Biometric Technology for Human Identification VI, vol.
7306. International Society for Optics and Photonics, 2009, p. 730628.
[10] Y. Kim and T. Moon, “Human detection and activity classification based
on micro-doppler signatures using deep convolutional neural networks,
IEEE geoscience and remote sensing letters, vol. 13, no. 1, pp. 8–12,
2016.
[11] M. S. Seyfio˘
glu, A. M. ¨
Ozbay˘
gglu, and S. Z. Gurbuz, “Deep con-
volutional autoencoder for radar-based classification of similar aided
and unaided human activities,IEEE Transactions on Aerospace and
Electronic Systems, 2018.
[12] B. Jokanovi ´
c and M. Amin, “Fall detection using deep learning in
range-doppler radars,” IEEE Transactions on Aerospace and Electronic
Systems, vol. 54, no. 1, pp. 180–189, 2018.
[13] B. Erol, M. G. Amin, and S. Z. Gurbuz, “Automatic data-driven
frequency-warped cepstral feature design for micro-doppler classifica-
tion,” IEEE Transactions on Aerospace and Electronic Systems, vol. 54,
no. 4, pp. 1724–1738, 2018.
[14] A.-K. Seifert, A. M. Zoubir, and M. G. Amin, “Radar classification of
human gait abnormality based on sum-of-harmonics analysis,” in Radar
Conference (RadarConf18), 2018 IEEE. IEEE, 2018, pp. 0940–0945.
[15] A.-K. Seifert, L. Sch¨
afer, M. G. Amin, and A. M. Zoubir, “Subspace
classification of human gait using radar micro-doppler signatures,” 26th
Eur. Signal Process. Conf. (EUSIPCO), 2018.
[16] H. T. Le, S. L. Phung, A. Bouzerdoum, and F. H. C. Tivive, “Human
motion classification with micro-doppler radar and bayesian-optimized
convolutional neural networks,” in 2018 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018,
pp. 2961–2965.
[17] G. Li, R. Zhang, M. Ritchie, and H. Griffiths, “Sparsity-driven micro-
doppler feature extraction for dynamic hand gesture recognition,” IEEE
Transactions on Aerospace and Electronic Systems, vol. 54, no. 2, pp.
655–665, 2018.
[18] Y. Kim and B. Toomajian, “Hand gesture recognition using micro-
doppler signatures with convolutional neural network,IEEE Access,
vol. 4, pp. 7125–7130, 2016.
[19] Z. Zhang, Z. Tian, and M. Zhou, “Latern: Dynamic continuous hand
gesture recognition using fmcw radar sensor,IEEE Sensors Journal,
vol. 18, no. 8, pp. 3278–3289, 2018.
[20] G. Li, S. Zhang, F. Fioranelli, and H. Griffiths, “Effect of sparsity-
aware time–frequency analysis on dynamic hand gesture classification
with radar micro-doppler signatures,” IET Radar, Sonar & Navigation,
vol. 12, no. 8, pp. 815–820, 2018.
[21] L. Yang and G. Li, “Sparsity aware dynamic gesture recognition using
radar sensors with angular diversity,” IET Radar, Sonar & Navigation,
vol. 12, no. 10, pp. 1114–1120, 2018.
[22] B. Jokanovi´
c and M. Amin, “Suitability of data representation domains
in expressing human motion radar signals,” IEEE Geoscience and
Remote Sensing Letters, vol. 14, no. 12, pp. 2370–2374, 2017.
[23] H. Mitchell, “Image similarity measures,” in Image Fusion. Springer,
2010, pp. 167–185.
[24] B. Erol, M. G. Amin, and B. Boashash, “Range-doppler radar sensor
fusion for fall detection,” in Radar Conference (RadarConf), 2017 IEEE.
IEEE, 2017, pp. 0819–0824.
[25] T. Cover and P. Hart, “Nearest neighbor pattern classification,IEEE
transactions on information theory, vol. 13, no. 1, pp. 21–27, 1967.
[26] J. Wang, P. Neskovic, and L. N. Cooper, “Improving nearest neighbor
rule with a simple adaptive distance measure,Pattern Recognition
Letters, vol. 28, no. 2, pp. 207–213, 2007.
[27] Y. Rubner and C. Tomasi, “The earth movers distance,” in Perceptual
Metrics for Image Database Navigation. Springer, 2001, pp. 13–28.
[28] M.-P. Dubuisson and A. K. Jain, “A modified hausdorff distance for
object matching,” in Proceedings of 12th international conference on
pattern recognition. IEEE, 1994, pp. 566–568.
[29] C. Cortes and V. Vapnik, “Support vector machine,Machine learning,
vol. 20, no. 3, pp. 273–297, 1995.
[30] I. Jolliffe, “Principal component analysis,” in International encyclopedia
of statistical science. Springer, 2011, pp. 1094–1096.
[31] T.-K. Kim, J. Kittler, and R. Cipolla, “Discriminative learning and
recognition of image set classes using canonical correlations,” IEEE
Transactions on Pattern Analysis and Machine Intelligence, vol. 29,
no. 6, pp. 1005–1018, 2007.
[32] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. D. Piatko, R. Silverman,
and A. Y. Wu, “An efficient k-means clustering algorithm: Analysis and
implementation,” IEEE Transactions on Pattern Analysis & Machine
Intelligence, no. 7, pp. 881–892, 2002.
ResearchGate has not been able to resolve any citations for this publication.
Conference Paper
Full-text available
Radar-based monitoring of human gait has become of increased interest with applications to security, sports biomechanics, and assisted living. Radar sensing offers contactless monitoring of human gait. It protects privacy and preserves a person's right to anonymity. Considering normal, pathological and assisted gait, we demonstrate the effectiveness of radar in discriminating different walking styles. By use of unsupervised feature extraction methods utilizing principal component analysis , we examine five gait classes using two different joint-variable signal representations, i.e., the spectrogram and the cadence-velocity diagram. Results obtained with experimental K-band radar data show that the choice of signal domain and adequate pre-processing are crucial for achieving high classification rates for all gait classes.
Article
Full-text available
We propose a hand gesture recognition technique using a convolutional neural network applied to radar echo I-Q plot trajectories. The proposed technique is demonstrated to accurately recognize six types of hand gesture for two participants. The system consists of a low-cost 2.4-GHz continuous-wave monostatic radar with a single antenna. Radar echo trajectories are converted to low-resolution images and used for the training and evaluation of the proposed technique. Results indicate that the proposed technique can recognize hand gestures with average accuracy exceeding 90%.
Article
Full-text available
Traffic speed meters are important legal measuring instruments specially used for traffic speed enforcement and must be tested and verified in the field every year using a vehicular mobile standard speed-measuring instrument to ensure speed-measuring performances. The non-contact optical speed sensor and the GPS speed sensor are the two most common types of standard speed-measuring instruments. The non-contact optical speed sensor requires extremely high installation accuracy, and its speed-measuring error is nonlinear and uncorrectable. The speed-measuring accuracy of the GPS speed sensor is rapidly reduced if the amount of received satellites is insufficient enough, which often occurs in urban high-rise regions, tunnels, and mountainous regions. In this paper, a new standard speed-measuring instrument using a dual-antenna Doppler radar sensor is proposed based on a tradeoff between the installation accuracy requirement and the usage region limitation, which has no specified requirements for its mounting distance and no limitation on usage regions and can automatically compensate for the effect of an inclined installation angle on its speed-measuring accuracy. Theoretical model analysis, simulated speed measurement results, and field experimental results compared with a GPS speed sensor with high accuracy showed that the dual-antenna Doppler radar sensor is effective and reliable as a new standard speed-measuring instrument.
Book
This text discusses the skills and abilities that air-traffic controllers need. Its approach is international as air-traffic control practices throughout the world have to be mutually compatible and agreed.
Article
In this study, two radar sensors with angular diversity are used to recognise dynamic hand gestures by analysing the sparse micro-Doppler signatures. The radar echoes are firstly mapped into the time-frequency domain through the Gaussianwindowed Fourier dictionary at each radar sensor. Then the sparse time-frequency features are extracted via the orthogonal matching pursuit algorithm. Finally, the sparse time-frequency features at two radar sensors are fused and input into the modified-Hausdorff-distance-based nearest neighbour classifier to achieve the dynamic hand gesture recognition. The experimental results based on the measured data under three different experimental scenes demonstrate that (i) the recognition accuracy can be improved by fusing the features extracted at two radar sensors when each radar sensor works well on its own; (ii) the recognition accuracy produced by feature fusion keeps satisfactory even if one of the radar sensors has poor performance, which means that the feature fusion can improve the robustness of the recognition system; and (iii) it would be more helpful if the line-of-sights of the two radar sensors are set to be orthogonal to each other.
Article
Dynamic hand gesture recognition is of great importance in human-computer interaction. In this study, the authors investigate the effect of sparsity-driven time-frequency analysis on hand gesture classification. The time-frequency spectrogram is first obtained by sparsity-driven time-frequency analysis. Then three empirical micro-Doppler features are extracted from the time-frequency spectrogram and a support vector machine is used to classify six kinds of dynamic hand gestures. The experimental results on measured data demonstrate that, compared to traditional time-frequency analysis techniques, sparsity-driven time-frequency analysis provides improved accuracy and robustness in dynamic hand gesture classification.
Article
In this letter, we propose two methods for personnel recognition and gait classification using deep convolutional neural networks (DCNNs) based on multistatic radar micro-Doppler signatures. Previous DCNN-based schemes have mainly focused on monostatic scenarios, whereas directional diversity offered by multistatic radar is exploited in this letter to improve classification accuracy. We first propose the voted monostatic DCNN (VMo-DCNN) method, which trains DCNNs on each receiver node separately and fuses the results by binary voting. By merging the fusion step into the network architecture, we further propose the multistatic DCNN (Mul-DCNN) method, which performs slightly better than VMo-DCNN. These methods are validated on real data measured with a 2.4-GHz multistatic radar system. Experimental results show that the Mul-DCNN achieves over 99% accuracy in armed/unarmed gait classification using only 20% training data and similar performance in two-class personnel recognition using 50% training data, which are higher than the accuracy obtained by performing DCNN on a single radar node.
Article
Recently, hand gesture recognition systems have become increasingly interesting to researchers in the field of human-computer interfaces. Real-world systems for human dynamic hand gesture recognition is challenging as: 1) the system must be robustness to various conditions; 2) there is a rich diversity in how people perform hand gestures, making hand gesture recognition difficult; 3) the system must detect and recognize hand gestures continuously using unsegmented input streams in order to avoid noticeable lag between performing a gesture and its classification. In this paper, to address these challenges, we present Latern, a novel system for dynamic continuous hand gesture recognition based on Frequency Modulated Continuous Wave (FMCW) radar sensor. The radar system does not depend on lighting, noise or atmospheric conditions. We employ a recurrent three-dimensional convolutional neural network to perform classification of dynamic hand gestures. To enhance the processing performance, Connectionist Temporal Classification (CTC) algorithm is used to train the network to predict class labels from inprogress gestures in unsegmented input streams. The experimental results show that Latern is able to achieve high recognition rates of 96%, which is higher than state-of-the-art hand gesture recognition systems. In addition, the conclusion in this work can be used for real-time hand gesture recognition system design. IEEE
Article
Radar-based activity recognition is a problem that has been of great interest due to applications such as border control and security, pedestrian identification for automotive safety, and remote health monitoring. This work seeks to show the efficacy of micro-Doppler analysis to distinguish even those gaits whose micro-Doppler signatures are not visually distinguishable. Moreover, a 3-layer, deep convolutional autoencoder (CAE) is proposed, which utilizes unsupervised pre-training to initialize the weights in the subsequent convolutional layers. This architecture is shown to be more effective than other deep learning architectures, such as convolutional neural networks (CNN) and autoencoders (AE), as well as conventional classifiers employing pre-defined features, such as support vector machines (SVM), random forest (RF) and extreme gradient boosting (Xgboost). Results show the performance of the proposed deep CAE yields a correct classification rate of 94.2% for micro-Doppler signatures of 12 different human activities measured indoors using a 4 GHz continuous wave radar - 17.3% improvement over SVM.