654

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 2, MARCH 2014

Robust Sensor Fusion of Unobtrusively Measured

Heart Rate
Tobias Wartzek, Christoph Bruser, Student Member, IEEE, Marian Walter, Senior Member, IEEE,
and Steffen Leonhardt, Senior Member, IEEE

AbstractContactless vital sign measurement technologies often have the drawback of severe motion artifacts and periods in
which no signal is available. However, using several identical or
physically different sensors, redundancy can be used to decrease
the error in noncontact heart rate estimation, while increasing the
time period during which reliable data are available. In this paper,
we show for the first time two major results in case of contactless heart rate measurements deduced from a capacitive ECG and
optical pulse signals. First, an artifact detection is an essential preprocessing step to allow a reliable fusion. Second, the robust but
computationally efficient median already provides good results;
however, using a Bayesian approach, and a short time estimation
of the variance, best results in terms of difference to reference
heart rate and temporal coverage can be achieved. In this paper,
six sensor signals were used and coverage increased from 090%
to 8094%, while the difference between the estimated heart rate
and the gold standard was less than 2 BPM.
Index TermsBayes, capacitive ECG (cECG), fusion, quality
index, robust, unobtrusive.

I. INTRODUCTION
NOBTRUSIVE vital sign monitoring technologies can
nowadays be integrated into many everyday objects. This
enables implementation of unsupervised automatic monitoring
and, therefore, preventive monitoring without the need for interaction, expert knowledge, or disturbing a persons daily routine. Most of these technologies do not require mechanical contact with the patient, e.g., in case of video analysis [1], or at
least do not need direct or galvanic contact with the skin, e.g.,
magnetic induction monitoring or the capacitive electrocardiogram (cECG) [2], [3]. Also, because these methods require no
preparation before a measurement, they are often called zeropreparation techniques.
Despite these advantages, these technologies tend to have two
main drawbacks related to contactless measurement:
1) severe motion artifacts, which may totally disrupt the signal of interest;

Manuscript received February 4, 2013; revised June 24, 2013; accepted July
17, 2013. Date of publication July 19, 2013; date of current version March 3,
2014.
The authors are with the Chair of Medical Information Technology,
RWTH Aachen University, 52062 Aachen, Germany (e-mail: wartzek@
hia.rwth-aachen.de; brueser@hia.rwth-aachen.de; walter@hia.rwth-aachen.de;
leonhardt@hia.rwth-aachen.de).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/JBHI.2013.2274211

2) specific conditions are required to measure a signal which

is not always present. For example, a camera requires a
line of sight, and cECG electrodes need an appropriate
position relative to the body. If these requirements are not
met, no signal is available from this particular sensor.
Such situations may preclude highly reliable calculation of
heart rate if, for example, only one sensor is used; this is particularly important in case of an unsupervised monitoring scenario.
A possible solution for these problems is to use several different sensors, since it is unlikely that artifacts and periods without
information will occur at the same time in each sensor [4]. However, the use of several sensors presents the problem of how to
select the best ones. To overcome these drawbacks, sensor fusion may be an appropriate method. If the fusion algorithm
is properly implemented, the temporal coverage during which
valid information is available will probably increase and, at the
same time, measurement error is decreased.
Several fusion approaches for standard monitoring (e.g., conventional ECG) are available. In 1997, details on an algorithm
for robust fusion of the heart rate were published [5], [6]; this
algorithm is based on the tests of consensus between sensor
measurements, the predicted value, and physiologic credibility,
leading to several hypotheses on the current state. A Kalman
filter was used to estimate the most likely state [5], [6].
A similar approach is used to assess the signal quality of the
ECG and arterial blood pressure, and tracks each signal with a
separate Kalman filter based on their individual signal qualities.
Fusion is performed by weighting each estimate of the Kalman
filter [7], [8]. Another method uses a so-called hybrid median,
i.e., a temporal median filter (former values from one sensor)
and a structural median filter (values from several sensors at the
same time) [9]. Compared to the Kalman filter approach, this
latter method has the advantage that no assumption needs to be
made about the underlying process characteristics and statistics
of the artifacts. Furthermore, from a computational point of
view, it is very simple. However, it remains unknown how this
method will perform in case of long-lasting and strong artifacts.
For the estimation of respiratory rate, a fusion concept using
only the best value (i.e., the most likely value) was recently
published [10].
The Bayesian fusion is often used in aircraft detection and
tracking; in this case, several sensors provide different probabilities for each detected aircraft (e.g., [11]). In this paper, a robust
Bayesian fusion of the peak-to-peak heart rate based on three
unobtrusive cECG signals and three unobtrusive optical pulse
signals is developed. It is shown that artifacts do not need to
occur simultaneously (e.g., motion occurs only in one sensor),
and hence, temporal coverage may be increased, while the dif-

2168-2194 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

WARTZEK et al.: ROBUST SENSOR FUSION OF UNOBTRUSIVELY MEASURED HEART RATE

Fig. 1.

Basic principle of the fusion process presented here.

Fig. 2. Measurement setup of the bed and mattress with the integrated contactless sensors.

ference between the fused heart rate estimate and the reference
heart rate is reduced.
II. MATERIALS AND METHODS
The principle of the fusion concept is shown in Fig. 1. The sensors measure a signal which contains information about the heart
rate: in this case a cECG and an optical pulse signal. Afterward,
each signal is processed separately. First, peak detection is performed which results in a vector pl = [pl1 , . . . , plk , . . . , plK ] of
peak locations in which l describes the index of the corresponding sensor. An artifact detection algorithm calculates a quality
index kl for each detected peak plk . This allows us to estimate
a peak-to-peak frequency f l (or heart rate HR, respectively)
with a corresponding quality index. Finally, all peak-to-peak
frequencies are fused in the last processing step.
Since classified frequency values instead of raw data or
features are fused, this fusion concept is called decision fusion [12]. Furthermore, the proposed method can be classified
according to a definition of Durrant-Whyte, as a complementary (i.e., the sensors are independent and can be combined to
increase completeness) and competing (i.e., all sensors measure
the same quantities which may result in increased accuracy)
fusion [13].
A. Measurement Setup and Preprocessing
The measurement setup is shown in Fig. 2 and consists of a
bed and a mattress with the integrated sensors. The sensors are
cECG sensors with an integrated optical sensor, as described
in [14]. A cECG sensor is similar to a standard ECG sensor
with the main difference that the coupling is capacitive rather
than conductive [3]. This allows us to measure through, e.g.,
clothes, which act as a dielectric layer within a capacitor. The
integrated optical sensor emits light through the clothes into the
body and measures the reflected signal strength. This fulfills two
functions: first, it detects motion at the electrodebody interface,

655

and second, it acts as a reflective photoplethysmography sensor

which measures peripheral blood pulse in the persons back.
Therefore, physical integration of these two sensors enables to
acquire (simultaneously) information about heart rate based on
two different physical principles.
As the position of a persons body (in relation to alignment
with sensors) is a priori unknown, 12 sensors are integrated
and distributed over the mattress and the three best coupled
electrodes are chosen. The principle involved in selecting the
best coupled electrodes is described in [15]. At this point, we
assume that three cECG signals and three optical signals are
acquired during one measurement. However, although the best
electrodes are chosen, it cannot be guaranteed that they really
measure a perfect signal (e.g., in case of a hollow-back). An
example of the different signal qualities will be shown in the
following.
Ten healthy volunteers participated in the study: seven males
and three females, mean age 27.9(21) years, height 1.8(1) m, and
body mass index 24.2(40) kg m2 . Participants lay on the bed
for about 5 min with the request to simulate typical movements
during sleep (e.g., small, strong movements) to produce motion
artifacts. A patient monitor (MP70, Philips, Bblingen, Germany)
was used to record a conductive ECG as reference.
For peak detection, an open-source ECG detector was used for
the cECG [16]. For the optical pulse signal, a simple maximum
value detector based on a SavitzkyGolay-Filter was used.
Although artifact detection is not a focus of this paper, it
is an essential preprocessing step. This is because long/severe
motion artifacts frequently occur, which complicate the achievement of reliable and robust fusion. For this paper, a previously
published multivariate artifact detection was adapted [4], which
calculates a quality index kl of each peak plk for each single
electrode signal. Performance is enhanced by using the robust
minimum covariance determinant estimator [17] rather than the
T 2 -algorithm. The resulting quality index is a continuous value
(range 01) that describes the probability that a detected peak
is an artifact ( = 0) or a true peak ( = 1). However, for the
fusion process described below, it is not important which specific artifact detection approach is used. The only requirement
is that a continuous probability value for each peak is available
and that the most severe artifacts are reliably discarded (in this
case, all frequency values fk with a quality index 0.4).
The effect of the artifact detection is exemplified in Fig. 3, in
which exemplary two cECG signals and one optical signal are
shown. All the detected peaks are marked with a blue square
and all remaining peaks after artifact detection are marked with
a red circle. It is clearly visible that, on the one hand, the signal
quality is very diverse and even time variant and, on the other
hand, although artifacts may occur approximately at the same
time, they probably have a different effect on the measurement
signal and consequently on the result of the peak detector. Here,
the optical signal seems to be almost undisturbed, but that is not
always the case.
Based on the detected peaks, it is possible to calculate a beatto-beat heart rate, which is shown in Fig. 4 for the same time
interval as in the previous figure (see Fig. 3). For comparison, the
heart rate based on the reference conductive ECG is also plotted

656

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 2, MARCH 2014

TABLE I
NOTATION IN CASE OF A BAYESIAN FUSION

B. Fusion Algorithms

Fig. 3. Two cECG and one optical signal as an example for the effect of
artifacts, the peak detector, and the artifact detection method.

In this study, a Bayesian-based fusion algorithm is presented

and its performance is compared with two computationally simple methods, i.e., the median and the best. The two latter
methods are straightforward, whereas the first is based on assumptions about the process.
1) Median of all Sensors: The first algorithm calculates the
median of all peak-to-peak frequencies. The median is computationally very simple but robust against outliers
fkfus = medianl (fkl ).

(1)

If only two signals exist (either a peak is discarded in the artifact

detection or the sensor does not measure any signal at all), the
median becomes the arithmetic mean.
2) Selection of the Best Sensor: This algorithm is also computationally simple, as it always chooses the best available sensor based on the previously calculated quality index
fkfus = fkl m a x ,

Fig. 4. Calculated beat-to-beat heart rate and the effect of the artifact detection
motivating the fusion process.

in each subplot (between 7 and 9 s the reference also contained

an artifact which was manually discarded). Due to the artifacts
and the false-detected peaks, false heart rates are calculated,
visible in the large deviation to the reference heart rate. These
large deviations are reduced at the cost of a reduced temporal
coverage of an available heart rate as time intervals are discarded
(indicated by grey shading). The time intervals in which no
information is available do not need to occur simultaneously.
Since the artifact detection is not perfect, some false heart rate
values remain, e.g., in the second subplot at approximately 7 s.
Although it seems that the heart rate based on the optical
signal is superior to the heart rate based on the other signals, it
needs to be mentioned that this is not always the case. Just to
the contrary, the heart rate of the optical signal is not as perfect
as the heart rate based on the cECG, as it is more difficult to
find the exact peak location of the slow varying optical signal
compared to the R-peak of the QRS-complex, which will also
be shown at the end of this paper.

with lm ax = arg max kl .

(2)

l=1...L

Bearing in mind that artifact detection methods are not always

perfect, we need to evaluate whether the sensor with the best
quality index is in fact always the best one to use.
3) Bayesian Fusion: Both previous methods have the disadvantage that they rely on an imperfect quality index. It is
probably advantageous to include the previous sensor measurements and the previous fused estimates in the fusion process
based on probabilities, rather than relying on an imperfect quality index. Therefore, the algorithm developed here does not use
the quality index as such, but considers it to be a preprocessing
step which discards the most severe artifacts.
The algorithm is based on [11] and has been adapted and
extended to allow an iterative and probability-based fusion of
contactless deduced heart rate. For heart rate deduction, the
following conventions about the symbols (as noted in Table I)
are defined.
The goal of the Bayesian fusion is to estimate the probability
of the real (but unknown) state f based on the last measurements
f lk . For the sake of simplicity, at this point, it is assumed that
only two sensors are used. The Bayesian fusion then provides
two probabilities about the unknown state f , given all previous
measurements
P (f = fk1 |f 1k f 2k )

(3)

P (f = fk2 |f 1k f 2k ).

(4)

To derive the final fusion algorithm and to use Bayes theorem, the previous sensor values are separated from the current

WARTZEK et al.: ROBUST SENSOR FUSION OF UNOBTRUSIVELY MEASURED HEART RATE

657

value
P (f |f 1k f 2k ) = P (f |fk1 fk2 , f 1k 1 f 2k 1 ).

(5)

Using Bayes theorem, the posterior probability follows to

p osterior

  
P (f |f 1k f 2k ) = P (f |fk1 fk2 , f 1k 1 f 2k 1 )
likeliho o d

prior






P (fk1 fk2 |f, f 1k 1 f 2k 1 ) P (f |f 1k 1 f 2k 1 )
=
. (6)
P (fk1 fk2 |f 1k 1 f 2k 1 )




Fig. 5. Two sensor outputs compared to a needed probability distribution to

assess the other sensor outputs.

norm alization

The fusion result of the previous step is the prior knowledge of

the current step, as indicated in (6).
Although all sensors measure the same source (i.e., the heart
beat), the artifacts and the detected peaks are, to some extent,
statistically independent. First, motion artifacts do not necessarily need to occur simultaneously and identically in every sensor,
and therefore, the detected peaks will vary between the signals. This was also already indicated in Fig. 3. Second, different
peak detectors are used for the cECG and for the optical signal.
Therefore, for simplification, it is assumed that the derived peakto-peak frequencies are statistically independent. This allows us
to reformulate
P (fk1 fk2 |f, f 1k 1 f 2k 1 ) = P (fk1 |f, f 1k 1 )P (fk2 |f, f 2k 1 ) (7)
which can be inserted into (6) to get
P (fk1 |f, f 1k 1 )P (fk2 |f, f 2k 1 )P (f |f 1k 1 f 2k 1 )
.
P (fk1 fk2 |f 1k 1 f 2k 1 )
(8)
Finally, with repeated application of Bayes theorem, the final
fusion algorithm (9) can be derived. This estimates the posterior
probability based on each sensors actual and previous probability, and the previously estimated fusion result. Equation (9)
can easily be expanded to accommodate more than two sensors.

P (f |f 1k f 2k ) =

P (f | f 1k f 2k ) =

P (f | f 1k ) P (fk1 | f 1k 1 )
P (f | f 1k 1 )

P (f |f 1k 1 f 2k 1 )
P (f |f 2k )P (fk2 |f 2k 1 )

P (f |f 2k 1 )
P (fk1 fk2 |f 1k 1 f 2k 1 )

P (f |f 1k )P (f |f 2k )P (f |f 1k 1 f 2k 1 )
.
P (f |f 1k 1 )P (f |f 2k 1 )

(9)

However, up to this point, (9) cannot be used for the estimation of heart rate. Each sensor, and its subsequent stages for
estimating the peak-to-peak frequency, provides only one discrete frequency fkl at each step k. In contrast, when a Bayesian
fusion is applied each sensor generally provides a probability
for all possible states (e.g., in the case of three aircrafts, each
sensor provides three probabilities [11]). This is also apparent
in (9), as the probability of all sensors is needed for each frequency. A graphical description of this circumstance is given in
Fig. 5. Only two discrete sensor values are available; however,
a probability distribution like P (f |f 1k ) for the first sensor is
needed to assess the other sensor values.

Fig. 6. Principle of a Bayesian fusion illustrated using a synthetic set of

signals.

As a solution for this problem, we propose to estimate the

missing information by assuming that the frequency values for
the last 10 s of each sensor can be described by a normal distribu l ) can be estimated using standard
tion, whose parameters (
l ,
methods. The period of 10 s was chosen empirically: if the period is too long, the standard estimation will be too large, and
if the period is too short, there will be insufficient samples to
estimate heart rate. However, the standard specification states
that the display of a heart rate monitor has to be updated at least
once during this period, i.e., a filter algorithm could in principle
use the last 10 s [18]. However, the assumption of a normal distribution is probably somewhat simplistic. Although heart rate
has been shown to have a normal distribution among a large
patient population [19], this is not necessarily the case for an
individual, sick patient. For a highly accurate prediction of heart
rate, a simple normal distribution is probably not sufficient [20].
However, we do not use the model solely for prediction but to
select the best sensors. In the case of modeling errors, these
errors are present in all the sensors. Since we use only relative
relationships of the single probabilities instead of the predicted
value, the impact of model errors is not very large.

658

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 2, MARCH 2014

Fig. 7. Difference between the estimated heart rate and the reference heart
rate using different fusion algorithms.

Even within a 10 s period, there are only a few frequency values. This may lead to underestimation of the standard deviation
(SD) which, in turn, has a negative effect on the fusion algorithm. In the case of a very small SD, the Bayesian algorithm
becomes locked in a tight normal distribution and no longer accurately follows changes in heart rate. For this reason, the upper
l
value
ol of the 99% confidence interval of the estimated SD
is used

n1
l
l

.
(10)

o =
2
0,005;n 1
The denominator in (10) refers to the 2 -distribution and n
denotes the number of frequency values.
Finally, it is possible to estimate the necessary probabilities
in (9) to allow for the Bayesian fusion

12
1
P (f |f lk ) =
e
2
ol

f f l
k
ol

2

(11)

In principle, it is possible to choose the frequency with the

highest probability. However, the results can be slightly improved if all frequencies with an estimated probability of 5%
are discarded. The estimated fused frequency fkfus is then the
median of the remaining frequencies
fkfus = medianl (fkl ), with the condition: P (f |f lk ) > 5%.
(12)
III. RESULTS
A. Bayesian Fusion of the Synthetic Data
The principle of the Bayesian fusion algorithm is shown
in Fig. 6 using the synthetic data. Three frequency values
f 1 , . . . , f 3 , their corresponding estimated probabilities
P (f 1 ), . . . , P (f 3 ), and the fusion result f fus are shown. If

Fig. 8. Temporal coverage in which information on heart rate is available

using different algorithms.

Fig. 9. Example of the fused heart rate compared to the reference of the same
time interval as in Figs. 3 and 4.

only one signal deviates (see samples 3050), or its variance

is larger than the others (see samples 80100), its estimated
probability P (f 1 ) becomes very small. If all signals show a
large step, the algorithm instantly follows this step and correctly
selects the new value as the fused estimate. If two signals show
a large change in step (see samples 110150), the algorithm
first estimates a higher probability to the single signal due to
the prior estimates. As a worst case scenario, the algorithm
requires n samples to switch to the correct two similar signals.
However, such a large change in step (as shown in Fig. 6) is
unlikely to occur in actual heart rate monitoring.
B. Fusion Results of the Measured Data
The effects and results of the different processing steps, artifact detection and fusion applied to actual measurements, are
shown in Figs. 7 and 8. The error in heart rate estimation, i.e.,
the difference between the estimated and reference heart rate
HR = HRref HR, is shown in Fig. 7. To rate the achieved
results, the estimated peak-to-peak heart rates of all signals
(three cECG and three optical signals) and of all ten persons
without any fusion method applied are plotted in the left column of Fig. 7 (a total of 15 497 data points). In case of fusion

WARTZEK et al.: ROBUST SENSOR FUSION OF UNOBTRUSIVELY MEASURED HEART RATE

and/or artifact detection, the amount of estimated heart rate data

is smaller due to the discarded values. The maximum allowed
error in heart rate estimation of 5 BPM as specified in the
ANSI standard [18], is also shown in Fig. 7.
The temporal coverage (as percentage of the entire measurement period) is shown in Fig. 8. In case of no fusion
(three cECG and three optical signals) 10 persons = 60 data
values are shown; when fusion is applied ten data values are
shown.
With regard to errors, the estimated heart rate will be of no
use in the absence of either artifact detection or fusion as it will
contain a considerable number of large errors. Even fusion with
the robust median but without artifact detection as preprocessing will fail to provide usable estimations of heart rate. This
emphasizes the need for an artifact detection step in contactless
vital sign monitoring. Due to the large number of severe artifacts, it is not possible to estimate the undistorted real value
in the fusion process.
Applying artifact detection considerably reduces error, even
if no further fusion is performed (see third column in Fig. 7).
However, a considerable number of errors still remain. These
can be further diminished by applying the median or Bayes
fusion, with Bayes algorithm outperforming the median algorithm. Many of the remaining errors in heart rate estimation are
due to errors in the reference heart rate, since even the reference ECG measurement was disturbed by the movements of the
test persons. However, compared with the reference data, 98%
of all the estimated heart rate values differed by less than 2
BPM; in most cases, this is sufficient for beat-to-beat heart rate
estimation without temporal smoothing and with no additional
calculations such as heart rate variability.
Due to the imperfect artifact detection, artifacts may acquire
an erroneously large quality index which leads to wrong fusion
results if the best sensor is selected before further validation has
taken place. Therefore, in case of imperfect artifact detection,
this method should not be used.
Besides errors in heart rate estimation, another important parameter is temporal coverage of the measurement period for
which heart rates are available. For the purpose of comparison,
Fig. 8 (extreme left) shows the calculated peak-to-peak heart
rates without artifact detection or fusion. Since no artifact detection is performed, most of the signals have a coverage of
around 100%. However, if a signal contains long periods in
which no peak is detected at all, its temporal coverage is less
than 100%. This is not a problem if several signals are available
(in this example, each measurement consists of six signals), as
one signal usually contains some information. This is obvious
in the second column of Fig. 8 as the coverage is 100% for all
ten test subjects in the case of median fusion. However, since no
artifact detection has been applied, the signal still contains large
errors in heart rate estimation (as already shown in the previous
figure).
If artifact detection is implemented, temporal coverage reduces because of the discarded peaks (see the third column
in Fig. 8). In case of highly corrupted signals with many artifacts and noise, artifact detection may discard all the detected
peaks resulting in a coverage of 0% (see the bottom of the third

659

TABLE II
EXAMPLE OF DETAILED RESULTS OF ONE MEASUREMENT ACQUIRED
WITH A BAYESIAN FUSION

column). This is important because if (hypothetically) only one

sensor was used and this happened to be the one with a coverage
of 0%, no monitoring would be possible. However, if fusion is
applied, the redundancy of the other sensors can be used. All
fusion methods in this paper achieve a coverage of 8090%.
Coverage of the Bayesian fusion is only slightly lower than the
others; however, bearing in mind the error in heart rate estimation, the Bayesian fusion clearly performs best, closely followed
by the median fusion.
As the aforementioned results provide an overall summary of
the ten measurements, Table II presents details of an example
of one representative measurement. For the signals cECG1
cECG3 and opt1opt3, artifact detection was applied, resulting
in a coverage of 078%. The mean error of the estimated heart
rate HR ranges from 0 to 1 BPM, while the optical signals
contain the larger deviation as already mentioned before. After
the Bayesian fusion, the coverage of the fused signal f fus increases to 86%, while maintaining a mean error in heart rate of
about 0 BPM.
The fusion result is also exemplary shown in a time plot of
the heart rate in Fig. 9, which covers the same time period as
in Figs. 3 and 4. It is clearly visible that the temporal coverage
is now similar to the reference, while at the same time no large
deviation of the fused heart rate is visible.
IV. CONCLUSION
Contactless measurement of vital signs by means of sensors
integrated into everyday objects has the major drawback of severe motion artifacts, or even no signal at all. Especially, if only
one sensor is used, alignment between the person and the sensor
is likely to be insufficient for reliable measurement.
Two important results are presented here. First, a robust artifact detection as a preprocessing step is shown to be essential
for successful signal processing, such as calculation of heart
rate and fusion of several sensors. Second, using a fusion algorithm based on a median or Bayesian algorithm, signals with no
information or with distorted intervals do not (or only slightly)
influence the fused estimates, and enable increased coverage
with minimal error. With the fusion algorithm a coverage of
8090% was achieved, while maintaining a (maximum) error in heart rate estimation as low as 2 BPM. Without fusion, estimated heart rate ranged from 80 to 40 BPM with a

660

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 18, NO. 2, MARCH 2014

temporal coverage of 090%. These results indicate that in

a real-life measurement scenario where contactless sensors
are integrated into everyday objects, it is advisable to implement more than one sensor together with an appropriate sensor
fusion.

REFERENCES
[1] L. Scalise, N. Bernacchia, I. Ercoli, and P. Marchionni, Heart rate measurement in neonatal patients using a webcamera, in Proc. IEEE Int.
Symp. Med. Meas. Appl. Proc.. Budapest, Ungarn: IEEE, May 1819,
2012, pp. 14.
[2] M. Steffen, A. Aleksandrowicz, and S. Leonhardt, Mobile noncontact
monitoring of heart and lung activity, IEEE Trans. Biomed. Circuits
Syst., vol. 1, no. 4, pp. 250257, Dec. 2007.
[3] Y. M. Chi, T.-P. Jung, and G. Cauwenberghs, Dry-contact and noncontact
biopotential electrodes: Methodological review, IEEE Rev. Biomed. Eng.,
vol. 3, pp. 106119, 2010. DOI: 10.1109/RBME.2010.2084078
[4] T. Wartzek, B. Eilebrecht, J. Lem, H.-J. Lindner, S. Leonhardt, and
M. Walter, ECG on the road: Robust and unobtrusive estimation of heart
rate, IEEE Trans. Biomed. Eng., vol. 58, no. 11, pp. 31123120, Nov.
2011.
[5] J. M. Feldman, M. H. Ebrahim, and I. Bar-Kana, Robust sensor fusion improves heart rate estimation: Clinical evaluation, J. Clin. Monit., vol. 13,
no. 6, pp. 379384, 1997.
[6] M. H. Ebrahim, J. M. Feldman, and I. Bar-Kana, A robust sensor fusion
method for heart rate estimation, J. Clin. Monit., vol. 13, no. 6, pp. 385
393, 1997.
[7] L. Tarassenko, L. Mason, and N. Townsend, Multi-sensor fusion for
robust computation of breathing rate, Electron. Lett., vol. 38, no. 22,
pp. 13141316, Oct. 2002.
[8] Q. Li, R. G. Mark, and G. D. Clifford, Robust heart rate estimation from
multiple asynchronous noisy sources using signal quality indices and a
Kalman filter, Physiol. Meas., vol. 29, no. 1, pp. 1532, 2008.
[9] P. Yang, G. A. Dumont, and J. M. Ansermino, Sensor fusion using a
hybrid median filter for artifact removal in intraoperative heart rate monitoring, J. Clin. Monit. Comput., vol. 23, no. 2, pp. 7583, 2009.
[10] C. Orphanidou, S. Fleming, S. Shah, and L. Tarassenko, Data fusion
for estimating respiratory rate from a single-lead ECG, Biomed. Signal
Process. Control, vol. 8, no. 1, pp. 98105, 2013.
[11] S. Challa and D. Koks, Bayesian and Dempster-Shafer fusion, Sadhana,
vol. 29, no. 2, pp. 145174, 2004.
[12] J. Llinas and D. L. Hall, An introduction to multi-sensor data fusion, in
Proc. IEEE Int. Symp. Circuits Syst., Buffalo, NY, May. 31Jun. 3, 1998,
pp. 537540.
[13] H. F. Durrant-Whyte, Sensor models and multisensor integration, Int. J.
Robot. Res., vol. 7, no. 6, pp. 97113, 1988.
[14] T. Wartzek, R. Elfring, A. Jansen, B. Eilebrecht, M. Walter, and S. Leonhardt, On the way to a cable free operating theater: An operating table
with integrated multimodal monitoring, presented at the Comput. Cardiol., Hangzhou, China, Sep. 1821, 2011
[15] T. Wartzek, H. Weber, M. Walter, B. Eilebrecht, and S. Leonhardt, Automatic electrode selection in unobtrusive capacitive ECG measurements,
presented at the 25th Int. Symp. Comput.-Based Med. Syst., Rome, Italy,
Jun. 2222, 2012
[16] P. S. Hamilton, Open Source ECG Analysis Software Documentation,
2002.
[17] S. Verboven and M. Hubert, LIBRA: A MATLAB library for robust
analysis, Chemometr. Intell. Lab. Syst., vol. 75, no. 2, pp. 127136, 2005.
[18] Association for the Advancement of Medical Instrumentation. (2002).
ANSI/AAMI EC13:2002: Cardiac monitors, heart rate meters, and
alarms, [Online]. Available:http://webstore.ansi.org
[19] L. Tarassenko, A. Hann, and D. Young, Integrated monitoring and analysis for early warning of patient deterioration, Brit. J. Anaesth., vol. 97,
no. 1, pp. 6468, 2006.
[20] R. Barbieri, A point-process model of human heartbeat intervals: New
definitions of heart rate and heart rate variability, AJP: Heart Circulat.
Physiol., vol. 288, no. 1, pp. H424H435, 2004.

Tobias Wartzek was born in Krefeld, Germany,

in 1982. He received the Dipl.-Ing. degree in
electrical engineering with focus on information and communication technology from RWTH
Aachen University, Aachen, Germany, in 2008,
where he is currently working toward the Dr.-Ing.
(Ph.D.) degree at the Chair of Medical Information
Technology.
He is also working as a Research Assistant at
RWTH Aachen University. His research interests include modeling physiological systems, signal processing, and unobtrusive sensor technologies for biomedical measurements.

Christoph Bruser
(S05) was born in Troisdorf,
Germany, in 1983. He received the Dipl.-Ing. degree
in computer engineering from the RWTH Aachen
University, Aachen, Germany, where he is currently
working toward the Dr.-Ing. (Ph.D.) degree at the
Chair of Medical Information Technology.
He is also working as a Research Assistant
at RWTH Aachen University. His research interests include biosignal processing and classification
as well as unobtrusive physiological measurement
techniques.

Marian Walter (M97SM13) was born in

Saarbrucken, Germany, in 1966. He studied electrical
engineering, with a specialization in control engineering, at Technical University of Darmstadt, Darmstadt,
Germany, and received the Dipl.-Ing. degree in 1995
and the Dr.-Ing. degree in 2002.
He has worked in medical engineering industry
for three years and was appointed a Senior Scientist
and the Deputy Head at the Philips Chair of Medical
Information Technology at RWTH Aachen University, Aachen, Germany, in 2004. His research interests include noncontact monitoring techniques, signal processing, and feedback
control in medicine.

Steffen Leonhardt (M95SM06) was born in

Frankfurt, Germany, in 1961. He received the M.S.
degree in computer engineering from SUNY at Buffalo, NY, USA, the Dipl.-Ing. degree in electrical
engineering and the Dr.-Ing. degree in control engineering from the Technical University of Darmstadt, Darmstadt, Germany, and the M.D. degree in
medicine from J. W. Goethe University, Frankfurt,
Germany.
He has five years of R&D management experience
in medical engineering industry and was appointed a
Full Professor and the Head of the Philips endowed Chair of Medical Information Technology at RWTH Aachen University, Aachen, Germany, in 2003.
His research interests include physiological measurement techniques, personal
health care systems, and feedback control systems in medicine.