Remote Heart Rate Determination in RGB Data

An Investigation using Independent Component Analysis and Adaptive Filtering

Christian Wiede, Julia Richter, Andr

e Apitzsch, Fajer KhairAldin and Gangolf Hirtz

Department of Electrical Engineering and Information Technology,

Chemnitz University of Technology, Reichenhainer Str. 70, 09126 Chemnitz, Germany

Keywords:

Heart Rate Detection, rPPG, Vital Parameters, Image Processing, Independent Component Analysis, Adaptive

Filtering.

Abstract:

An emerging topic in the ﬁeld of elderly care is the determination and tracking of vital parameters, such as

the heart rate. This parameter provides important information about a person’s current health status. Within

the last years, various research focussed on this topic. The recognition of vital parameters is increasingly

relevant for our ageing society. This paper presents a method to remotely determine the human heart rate with

a camera. At this point, we suggest to use independent component analysis (ICA) and adaptive ﬁltering for a

robust detection. In our processing chain, we used different image processing techniques, e. g. face detection,

and signal processing techniques, e. g. FFT and bandpass ﬁltering, in this study. An evaluation with several

probands, illuminations, frame rates and different heart rate levels showed that we could achieve a mean error

of 4.36 BPM, which corresponds to CAND of 94.45 %, and a speed of 35 fps.

1 INTRODUCTION

In a steadily ageing society, taking care for el-

derly plays a major role. In the last years, several

technical assistance systems have been developed to

help elderly people in their daily activities (Meinel

et al., 2015) and to call help in case of emergencies

(Wohlrab et al., 2015). These works show that it is

possible to detect a fall in the home environment and

inform relatives or caregivers. However, these sys-

tems only act after the emergency occurred. The goal

of our project is to detect a person’s current health

status and act pre-emptively in case of indications for

possible emergencies. Thus far, current assistant sys-

tems are not able to perform this. They are not able to

decide whether a person is sleeping in an armchair or

has suffered a circulatory collapse and is unconscious.

One possibility to overcome this problem is to de-

tect human vital parameters by means of optical sen-

sors. These so-called physiological parameters (e. g.

heart rate, breath rate and oxygen saturation) are de-

tectable with normal cameras due to the fact that the

spatial and temporal resolution of the cameras in-

creased in the last years. Especially the heart rate de-

tection drew attention in recent research (Poh et al.,

2010; van Gastel et al., 2014). Inspired by these

works, this paper presents a novel possibility to ro-

bustly detect the heart rate. We used ICA and an

adaptive ﬁltering to robustly detect the heart rate from

remote. Furthermore, we realised a real-time imple-

mentation, so that this method is suitable for elderly

care applications.

A high beneﬁt of this method is its contact-less

working mode. This method proved to be very con-

venient for patients, because they do not have to wear

any devices. In that way, effects such as skin irri-

tations and discomfort can be avoided. Measuring

the heart rate allows the detection of bradycardia and

tachycardia at a very early stage, so that emergencies

can be avoided. Besides helping elderly people, such

a system would be most beneﬁcial for detecting the

health status of neonatals to avoid sudden infant death

syndrome, for monitoring a driver’s well-being or for

triage in hospitals.

This paper is organised as follows. Firstly, there

will be an overview of related studies in this ﬁeld.

Secondly, we introduce our method for remote heart

rate detection. Thereupon, we present our results,

which is followed by a discussion. Finally, we sum-

marise our ﬁndings and give an outlook on further de-

velopments.

240

Wiede, C., Richter, J., Apitzsch, A., KhairAldin, F. and Hirtz, G.

Remote Hear t Rate Determination in RGB Data - An Investigation using Independent Component Analysis and Adaptive Filtering.

DOI: 10.5220/0005694002400246

In Proceedings of the 5th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2016), pages 240-246

ISBN: 978-989-758-173-1

2 RELATED WORK

Measuring the human heart rate is an active ﬁeld of

research in the last century. The oldest and most used

approach is the electrocardiography (ECG) developed

by Willem Einthoven in 1901. An electrocardiograph

measures the electric potential between two points on

the skin of a human body. The resulting character-

istic curves originates from the conduction system of

the heart. By simply counting the existing spikes, it

is possible to determine the human heart rate. It is the

gold standard even nowadays.

Another method evaluates the volumetric changes

of the tissue which are caused by the blood ﬂow. This

method is called plethysmography. The heart rate

can be measured by the variations of air pressure,

impedance or strain. In 1937, Hertzman and Speal-

man recognised the potential of an optical method that

later was called photoplethysmography (PPG) (Hertz-

man and Spealman, 1937). When light is transmitted

through a tissue, it changes its wavelength (modula-

tion) depending on the blood ﬂow. The crucial part is

the pulsatile fraction of the arterial blood ﬂow, which

is only one percent of the overall transmitted light.

With this method, thin body parts, such as ﬁngers or

earlobes, are penetrated by light emitted by a photo

diode (Allen, 2007). The light that was not absorbed

is measured on the other side of the body part. This

method is also called transmissive PPG, and it is of-

ten used for pulse oximetry. Next to the transmissive

PPG there is the reﬂectance PPG that measures the

reﬂected light that is emitted from the tissue. The

SNR of the reﬂective PPG is one dimension smaller

than the SNR of the transmissive PPG. However, the

reﬂective PPG has the disadvantage of being contact-

based. Moreover, it requires external light sources.

An alternative way is the remote photoplethys-

mography (rPPG), which is contact-less and does not

need any external light source. The basic concept

goes back to the year 2005 and was described by

Humphrey et al. (K. Humphreys and Ward, 2005).

In 2007, Garbey et al. showed a possibility to solve

this challenge with a thermal camera (Garbey et al.,

2007). The ﬁrst idea to measure the human heart rate

with rPPG in the visible light spectrum was published

by Verkruysse et al. in 2008 (Verkruysse et al., 2008).

They recorded videos of the human face with a RGB

camera and did not use any other light sources about

from daylight and normal artiﬁcial light. This ambi-

ent light could be considered as an additional source

of noise. The distance between camera and propositi

was 1-2 m. The propositi were instructed not to move

while sitting to avoid movement artefacts. A ROI was

selected in the persons’ faces and a spatial averaging

was applied to all three colour channels. They deter-

mined the heart rate in an image sequence by using

Fast Fourier Transform (FFT). The heart rate is repre-

sented by the change of illumination in face.

Although these publications have a basic function-

ality, they suffered from a low accuracy and artefacts

in the signal (e. g. moving artefacts). Poh et al. pro-

posed a more advanced method (Poh et al., 2010; Poh

et al., 2011). They used independent component anal-

ysis (ICA) as a blind source separation for the three

colour channels. This resulted in a more stable heart

rate determination that is robust against small motion

artefacts. Later, this general idea was improved by

van Gastel et al. (van Gastel et al., 2014) by us-

ing only the forehead region and by applying several

temporal ﬁlters. Instead of ICA, Lewandowska et al.

(Lewandowska et al., 2011) suggested to use prin-

cipal component analysis (PCA), which gives accu-

rate results while less computational power is needed.

In another publication, chrominance-based rPPG was

introduced. At this point, colour difference signals

were used to eliminate specular reﬂections on the skin

(de Haan and Jeanne, 2013). A different way to solve

the problem of rPPG is the Eulerian video magniﬁ-

cation (Wu et al., 2012; Rubinstein, 2014). With this

method, small movements in images can be visualised

to the human eye. He et al. showed how to detect

rPPG with this method (He et al., 2014). An alterna-

tive way that uses Newtonian reaction is presented by

Balakrishnan 2013 (Balakrishnan et al., 2013). They

detected certain interest points in the images, applied

PCA and determined small movements caused by the

cyclic blood movement. Other methods demonstrated

the possibility to make rPPG invariant against motion

as well (Li et al., 2014; Wang et al., 2015).

In 2015, van Gastel et al. proved that rPPG also

works in the near infrared spectrum (van Gastel et al.,

2015). They showed that the same methods are appli-

cable for these wavelengths.

On the basis of pulse detection, it is possible to ex-

tract other important medical parameters such as the

morphology of the signal or the heart rate variability

(HRV).

In our paper, we propose an approach to combine

ICA and adaptive ﬁltering for a robust heart rate de-

termination.

3 METHODS

3.1 System Overview

The following section describes the methods we ap-

plied in order to design a system for rPPG. Figure 1

shows the major steps. At ﬁrst, images are acquired.

Remote Heart Rate Determination in RGB Data - An Investigation using Independent Component Analysis and Adaptive Filtering

241

t=0

t=1

t=2

(1) Face Detection and

Forehead Extraction

(2) Temporal Signal Extraction

(3) Bandpass Filtering

(4) Independent Component Analysis

(5) Fast Fourier Transform

(6) Adaptive Filtering

Heart Rate

Figure 1: This overview shows the system functionality.

First of all, the face is detected and a forehead region is

deﬁned in the image (1). Then, three mean temporal sig-

nals are extracted from the RGB channels (2). This step

is followed by a bandpass ﬁlter (3). The ICA splits the

three colour channels into three independent components

(4). By using Fast Fourier Transform, existing frequencies

in these components are determined and the highest peak in

the spectrum is detected as the preliminary heart rate (5).

With an adaptive ﬁltering, quick changes in the signal can

be compensated (6), so that as ﬁnal result, the heart rate can

be obtained.

This step is followed by a face detection, the selection

of the forehead region and the extraction of the three

colour channels. A bandpass limits the frequencies to

the natural limits of the human heart rate. Afterwards,

an ICA is applied to the three channels. A frequency

analysis determines the ﬁnal heart rate. With the aid

of an adaptive ﬁltering, outliers can be eliminated.

A normal RGB camera (Basler ac640-100gc) was

used for image acquisition. It provides a VGA reso-

lution and alterable frame rates can be adjusted. For

each recording, the frame rate was set to a ﬁxed value

(30 fps or 50 fps) to have equidistant time steps for

the signal processing. The automatic controls for ex-

posure time and white balancing were switched off to

avoid invalid changes in the signal.

3.2 Face Detection and Forehead

Region Extraction

Due to the fact that our method works colour-based,

we need regions in the image where skin pixels are

visible. At this point, the face is the most suitable re-

gion. Normally, the face is not occluded by garments

and usually has a vertical orientation, which simpli-

ﬁes the detection.

In the last decades, face detection is a well-studied

topic. We used the Viola-Jones Detector (Viola and

Jones, 2004) in OpenCV, which classiﬁes Haar-like

features in cascades. The face detection itself gener-

ates bounding boxes with the face inside. There are

several face regions that are not suitable for heart rate

determination, because they do not provide signiﬁ-

cant information, e. g. hair and eyebrows, or because

they show strong movement artefacts, e. g. eyes and

mouth. Therefore, only a part of the face region was

taken into account. For a better signal quality, a re-

gion of interest (ROI) was selected that provides a rel-

atively low noise-affected heart rate signal. For this,

the forehead region has been chosen, since it disposes

sufﬁcient superﬁcial vessels because of the thin skin.

Moreover, it shows a good light reﬂection character-

istic with a low light absorption of the tissue. Further-

more, movement artefacts are signiﬁcantly less fre-

quent than in other face regions. The ROI is deﬁned

by a static window within the face region that has al-

ways the same position with respect to bounding box

of the face.

3.3 Temporal Signal Extraction and

Band Pass Filtering

After having selected the forehead region, all pixels

inside the ROI are summarised and averaged. These

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

242

0 1,000 2,000

−10

Time [Samples]

Magnitude

ch1(t)

0 1,000 2,000

−5

Time [Samples]

ch2(t)

0 1,000 2,000

−5

Time [Samples]

ch3(t)

Figure 2: In result of the ICA, there are three independent components. The third component ch3(t) showed the best result in

our implementation.

operations are performed for all three RGB colour

channels:

mean

) =

roi

∑

roi

∑

i, j

) (1a)

mean

) =

roi

∑

roi

∑

i, j

) (1b)

mean

) =

roi

∑

roi

∑

i, j

) (1c)

i, j

, G

i, j

and B

i, j

denote a single pixel in the ROI

of the corresponding channel. n

ROI

is the number of

all pixels in the ROI. Consequently, for every speciﬁc

moment in time t

, there is one value for each colour

channel that represents the mean value R

mean

, G

mean

and B

mean

In order to exclude implausible frequencies that

cannot represent the human heart rate, a bandpass ﬁl-

ter BP is applied, see Equation (2a) to Equation (2c).

Only frequencies higher than 0.5 Hz (30 BPM) and

lower than 3 Hz (180 BPM) are considered during

further computations, i. e. the signals R

, G

and

. For our practical implementation, we used a FIR

ﬁlter with 201 ﬁlter components. This ﬁlter has a lin-

ear phase response, which means that all frequencies

have the same group delay. Moreover, the ﬁltering

has the effect of a pre-whitening, which is necessary

for the ICA.

(t) = BP(t) ∗ R

mean

(t) (2a)

(t) = BP(t) ∗ G

mean

(t) (2b)

(t) = BP(t) ∗ B

mean

(t) (2c)

3.4 Independent Component Analysis

The three colour channels contain several sources of

image noise and artefacts, e. g. motion. The objective

is to ﬁnd the underlying, original signals and extract

the pulse signal by decomposing the colour channels.

One possibility for decomposition is the ICA. This

method assumes that our observations are a linear

combination of the independent sources. Hence, it

is called blind source separation. In the general equa-

tion~x = A

s,~x denotes the vector of the observed com-

ponents, A is the so-called mixing matrix with lin-

ear concatenated elements and

s represents the inde-

pendent source components. For this application, the

equation can be formulated as:





(t)









1,1

1,2

1,3

2,1

2,2

2,3

3,1

3,2

3,3









(t)





(3)

For this implementation, we used the FastICA ap-

proach of Hyv

arinen (Hyv

arinen, 1999). It is accu-

rate, fast and available for several programming lan-

guages.

In Figure 2, all three components are displayed.

The component with the highest periodicity in the sig-

nal, which is visible in the harmonics of the spec-

trum, is most likely the component we are looking

for. In our study, the third independent components

(t) shows the highest periodicity. In order to re-

move the high-frequent noise that was caused by the

ICA, we applied a smoothing window, which aver-

ages over three samples. This lowpass is denoted as

LP.

3smooth

= LP(t) ∗ ch

(t) (4)

The smoothed result ch

3smooth

is shown in

Figure 3.

Remote Heart Rate Determination in RGB Data - An Investigation using Independent Component Analysis and Adaptive Filtering

243

500

1,000

1,500

−5

Time [samples]

Magnitude

Figure 3: This plot shows the smoothed third independent

component.

3.5 Frequency Analysis

Finally, the frequencies of the signal ch

3smooth

are

determined using the Fast Fourier Transform (FFT).

Since the heart rate can change over time, we per-

formed the FFT in a small window of three seconds.

In this short interval, the human heart rate will not

change considerably. In frequency domain, the most

prominent peak is selected. This peak represents the

heart rate HR at time t

HR(t

) = max(

FFT(ch

3smooth

)

) (5)

500

1,000

1,500

BPM

Magnitude

Spectrum of the third smoothed component

Figure 4: The Fast Fourier Transform provides the fre-

quency spectrum of the third smoothed component. To in-

crease the frequency resolution, zero padding was applied.

The maximum peak is for this example at 72.5 BPM.

In order to increase the frequency resolution, we

applied zero padding to the signal. This guarantees a

more accurate quantisation without changing the sig-

nal information.

3.6 Adaptive Filtering

In order to eliminate false detections appearing in the

form of sudden changes in the heart rate from one

window to the next, we suggest the usage of an adap-

tive ﬁlter. If the absolute difference between the cur-

rent heart rate value HR(t

) and the previous heart rate

value HR(t

− ∆t) is higher than a threshold T) of 10

BPM, a sliding average is computed using the last 10

values according to Equation (6) and Equation (7). ∆t

denotes the time interval between to samples.

T =

HR(t

− ∆t) − HR(t

)

(6)

(

HR(t

) if T < 10 BPM

∑

t=t

−9

HR(t) otherwise.

(7)

4 EXPERIMENTAL RESULTS

AND DISCUSSION

4.1 Setting

For our experiments, we used a Basler RGB cam-

era (acA640-100gc). This camera records data with a

ﬁxed frame rate. Both automatic exposure time con-

trol and automatic white balancing are switched off.

As far as lighting sources are concerned, we tested

multiple subjects under different lighting conditions,

whereas the only lighting sources were daylight and

the illumination of the interior lights in the room. The

distance between camera and the probands was be-

tween one and two meters. The probands were in-

structed to sit still with the face directed to the cam-

era. The recorded sequences have a length between

30 seconds and 2 minutes.

As reference we used a Polar FT1 heart rate mon-

itor. This system is composed of a heart rate sen-

sor, which is mounted on a chest strap, and a display,

which shows the current heart rate. The display is

visible in the corresponding videos as well, with the

result that the current reference heart rate can be com-

pared with our results.

In order to generate more variability in the heart

rate, the probands were asked to perform squats be-

fore the recording of some sequences. This results in

higher heart rates and is more representative with re-

spect to the natural range of the human heart rate.

In the next sections the accuracy, the utilization

of other colour channels than RGB and the computa-

tional speed are considered and discussed.

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

244

4.2 Accuracy

In our recordings, the average heart rate ranged from

67 BPM to 114 BPM. The mean error is calculated

for each window from t

to t

end

as the absolute dif-

ference between the measured heart rate HR

(t) and

the reference heart rate HR

Re f

(t), see equation 8.

To measure the accuracy, we used this mean error

and the complement of the absolute normalised dif-

ference (CAND) as evaluation criterion, as shown in

equation 9.

Mean =

end

∑

t=t



Re f

(t) − HR

(t)



(8)

CAND = 1 −



Re f

− HR



Re f

(9)

In our publication, the CAND is expressed in per-

cent. The higher this value, the better is the algorithm.

In our overall results, the mean error of the heart rate

detection is 4.36 BPM and the CAND is 94.45 %.

This shows that we can accurately detect the human

heart rate with our method. For our addressed appli-

cation ﬁelds, these errors are reasonably small.

Table 1: Mean error and the CAND for the different image

sequences.

Sequence Mean Error in BPM CAND in %

Sequence 1 5.63 92.34

Sequence 2 2.59 96.27

Sequence 3 3.36 95.24

Sequence 4 5.54 94.77

Sequence 5 4.85 93.64

We demonstrated that the ICA and the adaptive

ﬁltering reduce the error signiﬁcantly. However, the

problem of ICA is the selection of the most suitable

component. Methods such as autocorrelation or spec-

tral density could solve this issue.

In another experiment, we tested our algorithm for

another colour space, the HSI. The HSI color space

splits the channels in hue, saturation and intensity.

We expected that the decoupling of intensity and hue

shows better result than RGB. However, the results

were a tenfold worse. That indicates that our pro-

posed method is not suitable for the HSI colour space.

4.3 Computational Speed

In order to increase the speed of our algorithm, we re-

implemented the Matlab algorithm in C++ by using

the computer vision library OpenCV. For the C++ im-

plementation, we measured the computational speed.

For the tests, an Intel core i7 quad core processor with

2.9 GHz and 16 GB RAM was used. For compiling

GCC 5.2 was used.

In the case of computing the ICA in every time

step, the mean processing time, starting from image

acquisition to the derivation of the current heart rate,

is 28 ms. This corresponds approximately to 35 fps,

which demonstrates that our method runs ﬂuently on

the described system. A speed comparison with other

publications was not possible, because this informa-

tion is not provided in literature.

A possibility to increase the speed is to calculate

the mixing matrix of the ICA only every n windows

or using the result of the ICA of the previous time step

for initialisation. Another acceleration option is to use

only R and G components of the RGB channel for

the ICA, because we assume that these two channels

contain the main part of the actual heart rate signal.

As a consequence, the computational effort will be

less with probably the same accuracy.

5 CONCLUSIONS

In this paper, we presented a method for remote heart

rate determination using ICA and adaptive ﬁltering.

With regard to applications in domestic environments

and for elderly care, the obtained results are adequate.

For other use cases, especially in clinical environ-

ments, where highly accurate measurements are re-

quired, accuracy has to be improved.

For future work, the algorithm should also be ro-

bust against motion artefacts. One solution could be a

feature tracking on the forehead region.

By porting the algorithm to an embedded system,

a more ﬂexible and praxis-oriented solution could be

achieved.

With the help of such a system, it could be possi-

ble to pre-emptively detect emergencies in domestic

environments.

ACKNOWLEDGEMENTS

This project is funded by the European Social Fund

(ESF). We would like to thank all probands who took

part in the experiments and supported us with their

video records.

REFERENCES

Allen, J. (2007). Photoplethysmography and its application

in clinical physiological measurement. Physiological

Measurement, 28(3):R1–R39.

Remote Heart Rate Determination in RGB Data - An Investigation using Independent Component Analysis and Adaptive Filtering

245

Balakrishnan, G., Durand, F., and Guttag, J. (2013). Detect-

ing pulse from head motions in video. In Computer

Vision and Pattern Recognition (CVPR), 2013 IEEE

Conference on, pages 3430–3437.

de Haan, G. and Jeanne, V. (2013). Robust pulse rate from

chrominance-based rppg. Biomedical Engineering,

IEEE Transactions on, 60(10):2878–2886.

Garbey, M., Sun, N., Merla, A., and Pavlidis, I. (2007).

Contact-free measurement of cardiac pulse based on

the analysis of thermal imagery. Biomedical Engi-

neering, IEEE Transactions on, 54(8):1418–1426.

He, X., Goubran, R., and Liu, X. (2014). Using eulerian

video magniﬁcation framework to measure pulse tran-

sit time. In Medical Measurements and Applications

(MeMeA), 2014 IEEE International Symposium on,

pages 1–4.

Hertzman, A. B. and Spealman, C. R. (1937). Observations

on the ﬁnger volume pulse recorded photoelectrically.

American Journal of Physiology, 119:334–335.

Hyv

arinen, A. (1999). Fast and robust ﬁxed-point algo-

rithms for independent component analysis. Neural

Networks, IEEE Transactions on, 10(3):626–634.

K. Humphreys, C. M. and Ward, T. (2005). A cmos camera-

based system for clinical photoplethysmographic ap-

plications. In Proceedings of SPIE, volume 5823,

pages 88–95.

Lewandowska, M., Ruminski, J., Kocejko, T., and Nowak,

J. (2011). Measuring pulse rate with a webcam - a

non-contact method for evaluating cardiac activity. In

Computer Science and Information Systems (FedC-

SIS), 2011 Federated Conference on, pages 405–410.

Li, X., Chen, J., Zhao, G., and Pietikainen, M. (2014).

Remote heart rate measurement from face videos un-

der realistic situations. In Computer Vision and Pat-

tern Recognition (CVPR), 2014 IEEE Conference on,

pages 4264–4271.

Meinel, L., Richter, J., Schmidt, R., Findeisen, M., and

Hirtz, G. (2015). Opdemiva: An integrated assis-

tance and information system for elderly with demen-

tia. In Consumer Electronics (ICCE), 2015 IEEE In-

ternational Conference on, pages 76–77.

Poh, M.-Z., McDuff, D., and Picard, R. (2010). Non-

contact, automated cardiac pulse measurements using

video imaging and blind source separation. Optics Ex-

press, 18(10):10762–10774.

Poh, M.-Z., McDuff, D., and Picard, R. (2011). Ad-

vancements in noncontact, multiparameter physiolog-

ical measurements using a webcam. Biomedical En-

gineering, IEEE Transactions on, 58(1):7–11.

Rubinstein, M. (2014). Analysis and Visualization of Tem-

poral Variations in Video. PhD thesis, Massachusetts

Institute of Technology.

van Gastel, M., Stuijk, S., and de Haan, G. (2015). Motion

robust remote-ppg in infrared. Biomedical Engineer-

ing, IEEE Transactions on, PP(99):1–1.

van Gastel, M., Zinger, S., Kemps, H., and de With, P.

(2014). e-health video system for performance anal-

ysis in heart revalidation cycling. In Consumer Elec-

tronics Berlin (ICCE-Berlin), 2014 IEEE Fourth In-

ternational Conference on, pages 31–35.

Verkruysse, W., Svaasand, L. O., and Nelson, J. S. (2008).

Observations on the ﬁnger volume pulse recorded

photoelectrically. Optics Express, 16(26):21434–

21445.

Viola, P. and Jones, M. J. (2004). Robust real-time face

detection. International Journal of Computer Vision,

57(2):137–154.

Wang, W., Stuijk, S., and de Haan, G. (2015). Exploiting

spatial redundancy of image sensor for motion robust

rppg. Biomedical Engineering, IEEE Transactions on,

62(2):415–425.

Wohlrab, D., Heß, M., Apitzsch, A., Langklotz, M.,

Schwarzenberger, A., Bilda, S., Schulz, H., Hirtz,

G., and Mehner, J. (2015). Hom-e-call - an en-

hanced fall detection system based on accelerometer

and optical sensors applicable in domestic environ-

ment. In Jaffray, D. A., editor, World Congress on

Medical Physics and Biomedical Engineering, June 7-

12, 2015, Toronto, Canada, volume 51 of IFMBE Pro-

ceedings, pages 1461–1464. Springer International

Publishing.

Wu, H.-Y., Rubinstein, M., Shih, E., Guttag, J., Durand, F.,

and Freeman, W. T. (2012). Eulerian video magniﬁ-

cation for revealing subtle changes in the world. ACM

Trans. Graph. (Proceedings SIGGRAPH 2012), 31(4).

ICPRAM 2016 - International Conference on Pattern Recognition Applications and Methods

246