You are on page 1of 4

DETECTION OF MULTIPLE HEARTBEATS USING DOPPLER RADAR †

Qin Zhou, Jianhan Liu, Anders Høst-Madsen, Olga Boric-Lubecke, Victor Lubecke
Department of Electrical Engineering, University of Hawaii
ABSTRACT small compared to the wavelength, a circuit that couples both
the transmitted and reflected waves to a mixer can produce
Doppler radar life sensing has shown promise in medical and an output signal with a low-frequency component that is
security applications. The current paper considers the directly proportional to the movement. This is the case when
problem of determining the number of persons in a given measuring chest surface motion related to respiration and
area (e.g., a room) using the Doppler shift due to heartbeat. heart activity. Figure 1 illustrates this concept. Internal body
The signal is weak and time-varying, and therefore poses a reflections are greatly attenuated (more severely with
complicated signal processing problem. We develop a increasing frequency) and will not be considered here.
generalized likelihood ratio test (GLRT) based on a model of The current paper focuses on the problem of detecting the
the heartbeat, and show that this can be used to distinguish
between the presence of 2, 1, or 0 subjects, even with a
single antenna. We further extend this to N antennas. The
results show that one can expect to detect up to 2N-1 subjects
using this technique.

1. INTRODUCTION

Doppler radar remote detection of heart and respiration


activity is a promising technique for unobtrusive health
monitoring and life sensing, with proof of concept
demonstrated for various applications [1-3]. However, so far
Fig. 1 A single antenna system with a single subject. A homodyne
this approach has been limited to sensing and detection of a radio is used to detect a phase shift proportional to chest
single subject. When there are two or more subjects present displacement due to cardiopulmonary activity.
in the environment, it is more challenging to isolate signals
from individual subjects. We are exploring separation of number of subjects within range of the device, and possibly
signals from multiple subjects in a single and multiple location, through direction of arrival (DOA). This has a
antenna systems. In a single antenna system, based on number of important applications: to find survivors in
expected different spectral signatures of individual disaster situations like earthquakes, and for military
cardiovascular related motion, it is possible to separate recognizance.
individual signals in frequency domain. In a multiple
antenna system (SIMO or MIMO), even if individual 2. SIGNAL MODEL
cardiovascular signatures are very similar, it is possible to
distinguish different subjects based on angle of arrival. In We assume a continuous wave (CW) radar system
this paper we will discuss theoretical background for both of transmits a single tone signal at frequency ω.. This signal is
these approaches, experimental results for a single antenna reflected from a target at a nominal distance d, with a time-
system, and simulation results for the SIMO system. varying displacement given by x(t). Suppose at first that the
The use of Doppler radar to sense vital signs is a signal from a given subject arrives from a single path. The
relatively recent innovation [1-3]. The clear advantage of sampled, passband received signal at the n-th antenna in a
such a method is that individuals can be monitored at a SIMO system with quadrature receivers can be written as
distance, with no contact required. By the Doppler effect, an [12]
RF wave reflected at a moving surface undergoes a rn (t ) = A exp( jω (t + x(t − nτ ) − nτ )) + wn (t ), n = 0..M − 1
frequency shift proportional to the surface velocity. If the
d
surface is moving periodically, such as the chest of person τ= sin(v)
breathing, this can be characterized as a phase shift c
proportional to the surface displacement. If the movement is


This research was sponsored by NSF under contract ECS0428975 and the U.S. army under contract W15P7T-05-C-B202
Here τ is of the order 1/(3⋅108), so that x(t-nτ)≈ x(t). After x i (t ) = ( Ai + α i (t )) cos(ω i t + ϕ i (t ) + θ )
mixing with the carrier wave, the basband received signal = Ai cos(ω i t + θ ) + n(t )
can be written as,
rn (t ) = A exp( j (ωx(t ) − nφ − θ )) + wn (t ), n = 0..M − 1
n(t ) = Ai (cos(ϕ i (t )) − 1) cos(ω i t + θ )

2πd + α i (t ) cos(ω i t + ϕ i (t ) + θ )
τ= sin(v) + ( Ai + α i (t )) sin(ω i t + θ ) sin(ϕ i (t ))
λ
for a linear array, with angle of incidence v and noise wn(t). If The amplitude variations αi(t) and ϕi(t) are zero mean
we collect the signal received at the M antennas into a vector, random processes modeling the HRV. We assume that these
this can be written as have zero mean not only at DC but at all frequencies, i.e.,
r (t ) = A exp( jωx(t ))s(φ ) + w (t ) T T

∫ α (t )e ∫ ϕ (t )e
1 − iωt 1 − iωt
(1) lim dt = lim dt = 0
s(φ ) = [1, exp( jφ ), K , exp( j ( M − 1)φ )]T T →∞ T
i
T →∞ T
i
−T −T
As ωx(t) is very small From this we can conclude that
exp( j (ωx(t )) ≈ (1 + jωx(t )) (2) T

∫ n(t )e
1 − iωt
is a good approximation, so that lim dt = 0
T →∞ T
r (t ) ≈ (1 + jωx(t ))s + w (t ) −T
Thus, the signal is characterized by a characteristic vector s. and therefore that
If there are now S subjects at different positions, they will ⎧1 ω = ω i
T

∫ x (t )e
1 − iω t
likely have different DOA vectors s, and we can write the lim dt = ⎨
⎩0 ω ≠ ω i
i
T →∞ T
total received signal as −T
S Thus, if we calculate the Fourier transform (and not the
r (t ) = ∑ exp( j(ωx (t ))ss s + b + w (t ) power spectral density) of the received signal, we can expect
s =1 to see peaks at the heartbeat frequencies ωi. The HRV is
S S

∑ ∑s
included in the term n(t) which can be seen as an additional
≈ jωx s (t )s s + s + b + w (t ) noise term.
s =1 s =1
where b is a DC component due to reflection from other 3. GLRT FOR DETECTION OF HEARBEAT
objects (e.g., walls, funiture) than humans. If we use the
approximation (2) we then get. This shows that the DC There exist a number of methods for determining model
component of the signal does not contain any information, order, e.g., AIC and MDL [9]. However, since our primary
and this can therefore be filtered out. objective is to determine the number of subject, we prefer to
state the problem as a tradional Neymann-Pearson hypothesis
2.1. Heartbeat signal model test [9].
Consider at first the case of a single receiver antenna with
The signal xi(t) generated by each subject consists of only the I-component available with at most 2 subjects in
respiration and heartbeat. The respiration is usually in the range. As argued in the previous section, we model the data
range 0-0.8 Hz and the heartbeat in the range 0.8-2 Hz. received in am interval as a mixture of two periodic signals,
While the respiration is a stronger signal than the heartbeat, y[k ] = A1 cos(ω1 kT ) + B1 sin(ω1 kT )
it is also more difficult to characterize and therefore to (1)
detect. In the current method we therefore remove most of + A2 cos(ω 2 kT ) + B 2 sin(ω 2 kT ) + n[k ]
the respiration by high pass filtering. The heartbeat signal where n[k] is white Gaussian noise (WGN) with power σ2,
itself is a rather complicated signal. It is nearly periodic, but and A1, A2, B1, B2, ω1, ω2, and σ2 are unknown. Since n[k]
B B

the period can vary from one beat to the next; this is called includes terms due to HRV, assuming it is WGN is indeed a
heart rate variability (HRV). HRV can be modeled as a rough approximation. However, in absence of detailed
random process [5] with strong periodicity. To simplify the information about HRV terms, one can argue from maximum
modeling, we filter the received signal with a bandpass filter entropy principles that a WGN model is the most robust. The
with a pass band of 0.8-2 Hz so that only the fundamental problem of determining if there are two or more or less than
frequency of the heartbeat is received. The resulting signal is two persons present can then be stated as the following
modeled as hypothesis test problem
H 1 : ( A1 , B1 ) ≠ (0,0), ( A2 , B2 ) ≠ (0,0)
H 0 : ( A2 , B2 ) = (0,0)
This is a composite hypothesis test problem with many
unknown parameters, and the best detector for this kind of
problem is commonly accepted to be the generalized
likelihood ratio test (GLRT) [12-13]. In the GLRT the t (y ) =
following test statistic is defined
∑ y[k ] − ( A cos(ω t ) − B sin(ω t ))s(φ )
N 2
max A , B , A , B ,ω ,ω ,σ 2 f (y ) min 1 1 1 1 1
k =1
t (y ) = 1 1 2 2 1 2
2
min ∑ y[k ] − ∑ ( A cos(ω t ) − B sin(ω t ) )s(φ )
max A , B , A =0, B =0,ω ,ω ,σ 2 f (y ) N 2
1 1 2 2 1 2
k =1 i =1 i i i i i
where f(y) is the likelihood function (probability density
function) for the received data y=[y[1],...,y[N]]. If t(y)>τ, Now the minimization ω1, ω2, φ1, φ2 is a non-linear problem
solved using a simple grid search.
where τ is some threshold, the GLRT decides H1 (two or
more persons), otherwise H0 (less than two persons). The Notice that the minimization with respect to φ1, φ2 gives
DOA as a by-product, so the methods can also be used to
threshold τ is determined so that a desired false alarm
localizing subjects.
probability is guaranteed. If H0 is decided, another GLRT can
then be used to decide between 0 or 1 subjects.
In the Gaussian case, the GLRT test statistic can be 4. EXPERIMENTAL SETUP
simplified to
0.2

t (y ) = 0.1

∑ ( y[k ] − A1 cos(ω1t ) − B1 sin(ω1t ) )


N 2
min 0
A1 , B1 ,ω1 k =1
-0.1
2

∑ ⎛ y[k ] −
∑ A cos(ω i t ) + Bi sin(ω i t ) ⎞⎟
N 2
min ⎜
k =1 ⎝
-0.2
ω ω
A1 , B1 , A2 , B2 , 1, 2
i =1 i ⎠ 0 1 2 3 4 5
time(sec)
6 7 8 9 10

The minimization over A1, A2, B1, B2 is a linear problem, but


B B

0.4
the minimization over ω1, ω2 is a non-linear problem, which is
0.2
currently solved using a simple grid search.
0

-0.2
3.1. Extension to multiple sensors
-0.4
0 1 2 3 4 5 6 7 8 9 10
time(sec)
Assume that there are multiple receiver antennas with both I
and Q-components, and that multipath is negligible. The
received signal can then be modeled by Fig.3. Top: wireless signal; Bottom: reference signal.
y[k ] = ( A1 cos(ω1 kT ) + B1 sin(ω1 kT ) )s(φ1 )
+ ( A2 cos(ω 2 kT ) + B 2 sin(ω 2 kT ) )s(φ 2 ) + n[k ] A single antenna system, with one subject in the antenna
field of view, is illustrated in Fig. 2. The return signal is
Where s(φ) is given by (1). The GLRT test statistic is is now
mixed with a sample of the transmitted signal to produce an
output voltage with its magnitude proportional to the phase
shift between them, which in turn is proportional to chest
displacement due to cardiopulmonary activity. All radar
measurements were conducted with the CW signal source at
2.4 GHz with 0 dBm output power.
Fig. 3 shows a typical heart beat signal, compared with a
reference signal measuring the pulse from a finger sensor.
The signal has been filtered with a lowpass filter with cutoff
10Hz. It can be seen that the signal, compared with the
reference signal, is very noisy, and the heartbeat cannot be
find simply from peak detection.

5. RESULTS

Fig. 4 shows the test statistic (2) applied to three different


set of measurements with a single antenna. The
measurements are at first filtered with a bandpass filtered
with a passband 0.8-2Hz to remove respiration and higher

Fig. 2. Photograph of an Experimental Doppler radar set-up.


order harmonics, and are then divided into (overlapping)
intervals of length 15s. This ensures that the model (1) is 6. CONCLUSION AND FUTURE WORK
reasonably accurate. The test statistic is now evaluated in
each interval. It shows that it is objectively possible to We have shown that it is possible to determine the
determine the number of persons, e.g., by using threshold of number of subjects within range using hypothesis testing,
1.25. Notice that once it has been determined that less than although we are dealing with very weak and complicated
two persons are present with the above test, another GLRT signals. The method has potential of detecting up to 2N
can then be used to distinguish 0 and 1 persons. subjects with N antennas. The performance is acceptable, but
Results do not always turn out as well as in Fig. 4. Test could possibly be improved by taking into account more
subjects with strong HRV is sometimes not detected. To characteristics of the signals. On the other hand, already the
remedy this either the SNR has to be improved, or multiple current method is very computationally complex. It can be
simplified by using an approximate minimization, for
1.9
example by using 2D FFT and peak search.
Two persons data If there is heavy multipath the current method cannot be
one person data
1.8
no person data
used. Essentially, the method used is a spatio-temporal DOA.
1.7 With heavy multipath blind source separation methods must
be used, for example methods inspired by space-time music
1.6
[6-8].
Test Statistics

1.5 Finally, we are also working on extracting heart rate and


1.4
HRV from the data, but this is beyond the scope of the
current paper.
1.3

1.2 REFERENCES
1.1
[1] J. C. Lin, “Microwave sensing of physiological movement and
1
0 20 40 60 80 100 120 140 160 180
volume change: a review,” Bioelectromagnetics, vol. 13, pp. 557-
Time(Sec) 565, 1992.
[2] K. M. Chen, D. Mirsa, H. Wang, H. R. Chuang, E. Postow,
Fig.4. GLRT test for 2/1/0 person with 1Ant
“An X-band microwave life detection system,” IEEE Transactions
of Biomedical Eng., vol. 33, pp. 697-70, 1986.
antennas applied.
We do not yet have a working experimental setup with [3] V.M. Lubecke, O. Boric-Lubecke, G. Awater, P.-W. Ong, P.
Gammel, R.-H. Yan, J.C. Lin, “Remote sensing of vital signs with
multiple antennas. Multiple antenna data is therefore telecommunications signals,” World Congress on Medical Physics
obtained from partial simulation. Reference signals are and Biomedical Engineering (WC2000), Chicago IL, July 2000.
measured for different persons. They are then multiplied with [4] D. Samardzija, O. Boric-Lubecke, A. Host-Madsen, V. M.
DOA vectors, and independent noise is added at each Lubecke, T. Sizer, A. D. Droitcour and G. T. A. Kovacs,
antenna. Fig. 5 shows some results. In this case we used test “Applications of MIMO Techniques to Sensing of Cardiopulmonary
subject with strong HRV. With a single antenna it is Activity”, IEEE/ACES International Conference on Wireless
therefore difficult to detect reliably if there are 1 or 2 Communications and Applied Computational Electromagnetics,
subjects. With more antennas it is clearly possible to 2005.
distinguish between these situations. [5] McSharry PE, Clifford GD, Tarassenko L, Smith L. “A
dynamical model for generating synthetic electrocardiogram
signals”. IEEE Trans. Biomed. Eng., vol 50(3), pp: 289-294, Mar.
1.7 2003.
1Ant,2p
1Ant,1p [6] A. Belouchrani, K. A. Meraim, J.-F. Cardoso, and E. Moulines,
1.6 1Ant,0p
4Ant,2p
“A blind source separation technique using second order statistics,”
4Ant,1p IEEE Trans. Signal Processing, vol. 45, pp. 434–444, Feb. 1997.
1.5
4Ant,0p [7] A. Belouchrani and M. G. Amin, “Time-frequency MUSIC,”
GLRT STATISTICS

1.4
IEEE Signal Processing Lett., vol. 6, pp. 109–110, May 1999.
[8] Bousbia-Salah, A.; Belouchrani, A.; Cichocki, A.;”Application
1.3 of time-frequency distributions to the independent component
analysis of ECG signals”. Signal Processing and its Applications,
1.2 Sixth International, Symposium on. 2001. vol.1, pp.238 – 241, Aug.
2001.
1.1 [9] S.M. Kay: “Fundamentals of statistical signal processing,” Vol.
2, Prentice Hall, 1998.
1
0 10 20 30 40 50 60 70
Time(Sec)

Fig.5. 1 Ant versus 4 Ant GLRT test