Improved Pulse Detection from Head Motions using DCT

Ramin Irani, Kamal Nasrollahi and Thomas B. Moeslund

Visual Analysis of People Laboratory, Aalborg University (AAU), Sofiendalsvej 11, 9200 Aalborg, Denmark

Keywords: Heartbeat Rate, Head Motion Detection, Trajectory, Feature Point Tracker, Principle Component Analysis,

Discrete Cosine Transforms, Electrocardiogram.

Abstract: The heart pulsation sends out the blood throughout the body. The rate in which the heart performs this vital

task, heartbeat rate, is of curial importance to the body. Therefore, measuring heartbeat rate, a.k.a. pulse

detection, is very important in many applications, especially the medical ones. To measure it, physicians

traditionally, either sense the pulsations of some blood vessels or install some sensors on the body. In either

case, there is a need for a physical contact between the sensor and the body to obtain the heartbeat rate. This

might not be always feasible, for example, for applications like remote patient monitoring. In such cases,

contactless sensors, mostly based on computer vision techniques, are emerging as interesting alternatives.

This paper proposes such a system, in which the heartbeats (pulses) are detected by subtle motions that

appear on the face due to blood circulation. The proposed system has been tested in different facial

expressions. The experimental results show that the proposed system is correct and robust and outperforms

state-of-the-art.

1 INTRODUCTION

Heartbeat rate is obviously a vital sign of human

body's activity and its measurement is of great

importance in many applications, for instance,

fitness assessment, training programs and medical

diagnosis. For example, in fitness assessment during

the exercise, heartbeat rate is used as a crucial sign

that helps to assess the condition of cardiovascular

system. Here it can be used also for ensuring the

safety of the process. If the heartbeat rate goes

beyond the normal range, continuing the exercise is

not safe any longer.

Heartbeat rate is usually measured by devices

that take samples of heartbeats and compute the

beats per minute (bpm). Currently, one of the

popular non-invasive and standard devices for

measuring the heartbeat rate is electrocardiogram

(ECG). They are very accurate, but expensive. These

devices are electrode-based and therefore require

wearing adhesive gel patches or chest straps that

may cause skin irritation and slight pain.

Commercial pulse oximetry sensor is another

technique that is placed on specific parts of body

like fingertips or earlobe.

Though the above mentioned devices are

accurate, they are inconvenient as they need to have

physical contact with patient's body. Therefore,

developing contactless methods, which are based on

the patient's physiological signals, have recently

been considered as an interesting alternative for

measuring heartbeat rate. This technology would

also decrease the amount of cabling and clutter

related to Intensive Care Unit (ICU) monitoring,

long-term epilepsy monitoring, sleep studies, and

any continues heartbeat rate measurement (Poh,

2010). These contactless methods that are usually

based on computer vision techniques can be divided

into two groups. In the first group, known as

photoplethysmography (PPG) methods, usually a

red, or an infrared light is transmitted on the patients

(face or body) and the reflected light is sensed by the

system. The variations in the transmitted and the

reflected lights are then used to measure heartbeat

rate. Besides using dedicated light sources, the main

drawbacks of PPG systems are that they are

susceptible to motion artefact (Verkruysse, 2008);

(Humphreys, 2007); (Takano, 2007); (Hu, 2008);

(Wieringa, 2005).

In the second group of computer vision based

methods there is no need for a dedicated light

source. These methods assume that the periodic

circulation of the blood by the heart to the rest of the

body, including the head, generates some periodic

118

Irani R., Nasrollahi K. and Moeslund T..

Improved Pulse Detection from Head Motions using DCT.

DOI: 10.5220/0004669001180124

In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 118-124

ISBN: 978-989-758-009-3

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

subtle changes to the skin color of the face and also

generates some subtle head motions. These motions

are not usually visible to naked eyes but they can be

viewed by techniques like, for example, Eulerian

video magnification (Wu, 2012). These periodic

changes to the skin colors and head motions are then

utilized to measure heartbeat rate. For example, in

(Poh, 2010) periodic changes in the skin color of the

face has been used for this purpose. In this system

(Poh, 2010) face image of the subject is first found,

by a simple camera. Then, it is separated into its

colour channels and each channel is tracked

independently. For each of these tracked colour

channels, a trajectory is found. Then, all the

trajectories are fed to an Independent Component

Analysis (ICA) algorithm. The output of ICA,

presents independents sources that have caused

changes to the skin colour of the face. Then, it is

assumed that the most periodic output of ICA should

be generated by the most periodic source that is

present on the face, i.e., heartbeat. This system is

effective, but it suffers from sensitivity to skin color

and noise. It means, if the skin is not detected

properly, or if the captured facial video is noisy, the

system does not provide accurate results.

To overcome the sensitivity to noise and skin

detection of system (Poh, 2010), very recently in

(Balakrishnan, 2013) a motion-based contactless

system for measuring heartbeat rate was introduced.

As mentioned above, this method is based on the

fact that periodic circulation of the blood from the

heart to the body, including the head through the

aorta and carotid arteries, causes the head to move in

a cyclic motion (Wu, 2012). Similar to (Poh, 2010),

this system also uses a simple camera for recording

facial images of patients. Having detected the face,

they extracted vertical component of head motion by

tracking feature points, and generate some

trajectories for each feature point. These trajectories

are then filtered by a Butterworth filter to remove

the irrelevant frequencies. Next on the contrary

(Poh, 2010) they use Principle Component Analysis

(PCA) (instead of ICA) to decompose the filtered

trajectories into a set of source signals. Then, they

use the same assumption as (Poh, 2010), that the

most periodic signal is generated by the most

periodic source of the motion that is present in the

face, i.e., by heartbeat. To find the periodicity of the

outputs of PCA, they apply Fast Fourier Transform

(FFT) to the trajectories, and use the percentage of

total spectral power of the signal accounted for by

the frequency with the maximal power and its first

harmonic (Balakrishnan, 2013).

This method gives reasonable results when the

face is frontal and does not move. Our experiment

shows that involuntary motion and facial expression

causes dramatic effect on the accuracy of this

system. Furthermore, as mentioned above, this

system is based on using the frequency with

maximal power as the first harmonic of the

estimated heartbeat rate. But, this assumption is not

always true, especially when the facial expression is

changing. The proposed system in this paper

improves the system of (Balakrishnan, 2013) by

replacing the FFT with a Discrete Cosine Transform

(DCT). Furthermore, we show that involving a

moving average filter before the Butterworth filter

improves the results. It is shown that the proposed

system outperforms the system of (Balakrishnan,

2013), significantly.

The rest of this paper is organized as follows:

The clear problem statement and the contributions of

the proposed system are given in the next section.

Section 3 explains the employed methodology of the

proposed system. The experimental results are

reported in Section 4. Finally, the paper is concluded

in Section 5.

2 PROBLEM STATEMENT AND

MAIN CONTRIBUTION

The proposed system in this paper develops a vision-

based contactless algorithm for heartbeat rate

measurement using the assumption that periodic

blood circulation by the heart to the head generates

subtle periodic motion on the face. The proposed

system is based on the very recent work of

(Balakrishnan, 2013), but it advances this work by:

1) Replacing the FFT of the system of

(Balakrishnan, 2013) by a DCT, and

2) Using a moving average filter before the

Butterworth filter that is employed in

(Balakrishnan, 2013).

The proposed modifications are simple, but are

shown to be very effective. The results of the

proposed system are:

1) More correct compared to the results of the

system of (Balakrishnan, 2013) when they are

compared to the ground truth data.

2) More robust than the results of the system of

(Balakrishnan, 2013) when the face is moving or

facial expression is changing.

ImprovedPulseDetectionfromHeadMotionsusingDCT

119

Figure 1: The block diagram of the proposed system.

3 METHODOLOGY

The block diagram of the proposed system is shown

in Figure 1. As it can be seen from this figure, the

subject is continuously filmed by a Logitech

webcam with a resolution of 640x480 pixels. Then,

the subject's face is detected by Viola and Jones

(Viola, 2001) face detector. From the detected faces,

the regions of interest of our system, and

consequently the feature points are extracted and

tracked by the Lucas Kanade’s algorithm (Bouguet,

2000). Then, a moving average filter and a band pass

filter are applied to the vertical component of the

trajectories of each feature point to remove

extraneous frequencies and involuntary head

motions. Then, the filtered trajectories are fed to

PCA to find the strongest independent components.

Among these components, the most periodic one

belongs to heartbeat. To find this most periodic one,

we apply DCT to all the components obtained by

PCA. Each of these sub-blocks is explained in the

following subsections.

3.1 Face Detection

Locating the face in the scene refers to identifying a

region containing a human face. Viola and Jones

algorithm (Viola, 2001) has been employed for this

purpose which is based on Haar-like rectangular

features that are extracted from integral images. This

detector is fast and efficient, but it fails to detect

rotated faces and those which are of poor quality.

However, it works fine for the purposes of the

proposed system.

The regions detected by the Viola and Jones

detector cannot be directly used in our system, as it

contains the areas of eyes and the mouth which are

not good for the purposes of our system. Because

these areas are the most changeable areas of the

face, and they may change very much by any

changes in facial expression, eye blinking, etc.

Therefore, the trajectories obtained from these

changeable will not reflect the motion caused by

heartbeat. Instead, they reflect the motion caused by

the changes in their own positions due to the

changes in the facial expression. Tracking these

sensitive regions therefore does not produce stable

results. The most stable parts of the face, which are

robust against changes in the facial expressions, are

the forehead and the area around the nose. To keep

these regions, we first keep 50% (experimentally

obtained) of the width and 70% (experimentally

obtained) of the height of the region that is detected

the Viola and Jones’s face detector. Then, in this

refined region we remove the area of the eyes, by

removing all the pixels that are located in the range

of 25% to 45% (experimentally obtained) of the

height of the refined region (Figure 2).

Figure 2: The yellow box is returned by the Viola and

Jones face detector and the red boxes are those that are of

the interest of the proposed system.

3.2 Feature Points Selection

Having detected the regions of interest in the

previous sub-block of the system, in this step they

are fed to the Good Feature Tracking algorithm of

(Shi, 1994) to select the feature points. This

algorithm is based on finding and tracking the

corners. To do so, it calculates the minimal

eigenvalue of every point in our previously kept

regions of the face and rejecting corners with

minimal eigenvalues. Then, it goes through the

strongest corners and removes those features that are

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

120

too close the stronger features (Shi, 1994). To

increase the efficiency of this system, it is suggested

in (Balakrishnan, 2013) to divide the sub-regions

obtained from Viola and Jones detector into smaller

areas to achieve uniform selected regions. Therefore,

we have adopted this idea here.

3.3 Trajectory Generation and

Smoothing

To extract the motion trajectory signals from the

selected feature points in the previous subsection,

we have used Lucas Kanade’s algorithm (Bouguet,

2000) to obtain x and y components of feature points

inside our previously extracted regions of interest in

each frame. Since the very tiny motions of the head,

which are the basis for calculating the heartbeat rate

in this work, are due to the blood circulation through

aorta towards head (obviously in a vertical

direction), we only consider the y components of the

trajectories of the feature points in each frame.

The head motions are not only due to heartbeats

(transferred to the head by aorta), but may appear for

several reasons, for example, respiration, vestibular

activity, facial expression, speaking and so on. To

decrease the effects of the other sources, which

cause quite large motions, a moving average filter is

applied to the trajectories to smooth it (Figure 3).

This will be further explained in the experimental

results.

Figure 3: The effect of the employed moving average filter

on the y components of the trajectory of one of the tracked

feature points of one of the test subjects. The red and the

blue signals are the original and the filtered signals, on the

x axis of the above graph is the time and on the y axis is

the y position of the tracked feature point over time.

Then, to remove the irrelevant frequencies (any

frequency which might not be generated by the

heartbeat) a pass band filter (an 8

order

Butterworth filter) with cutoff frequency interval of

[0.75 5] Hz has been applied to the obtained

trajectory (Balakrishnan, 2013).

3.4 Signal Estimation

As mentioned above, the head motions are orginated

from different sources and only the one casued by

the blood circulation through aorta is reflecting the

heartbeat rate. To separate the sources of head

motions, we have applied a PCA algorithm to the

obtained trajectories. PCA converts the given

trajectories into a set of linearlly uncorrelated basis,

i.e., the principal components.

Having separated the sources using PCA, the

next step is to find the signal that has been generated

by the heartbeat. Following (Balakrishnan, 2013)

such a signal will be the most periodic signal. To

quantify the signal periodicity we have utilized DCT

as opposed to the system of (Balakrishnan, 2013)

which have used FFT. Having applied DCT, we only

keep those DCT components that carry the most

significant power of the signal. To do so, we use the

following algorithm:

 For the trajectory of the ith feature points,

i∈[1..N], S

o Calculate the DCT of the ith trajectory and

obtain SC

o Determine {K

}

which is the set of indexes

for {S

(t)} such that K

is the index of the M

first highest power components into SC

which consists 50% of power of S

 j∈[1..M

]

is number of components

which carry 50%of total power of S

)

o Determine {Kh

}

which is the set of the first 5

smallest index into {K

}

for each S

such that

2×Kh

be found on SC

 l = 1:5.

o The periodicity of the signal can be obtained

by:



















,2







/





 S

with largest Q

is the heartbeat rate signal, and

the heartbeat rate can be obtained as: FFT(IDCT

(min{Kh

})) × 60 bpm

The effect of the above DCT-based algorithm for

finding the heartbeat rate and its advantage over the

FFT of (Balakrishnan, 2013) has been shown in the

experimental results.

4 EXPERIMENTAL RESULTS

The proposed approach has been implemented in

Matlab R2013a. To be able to compare our system

ImprovedPulseDetectionfromHeadMotionsusingDCT

121

against state-of-the-art Balakrishnan et al.’s work

(Balakrishnan, 2013) we have recorded the actual

heartbeat rates of the test subjects by a Shimmer

wireless ECG (Electrocardiogram) sensor. This

sensor records and sends the ECG signals, to a

remote computer as a data file. Figure 4 (top) shows

a typical data that has been captured by this sensor.

The FFT of this signal is shown in Figure 4

(bottom). It can be seen from this figure, that the

FFT has 4 peaks on the frequencies 1.08, 2.14, 3.13,

and 4.22. These show that most of the power of the

recorded heartbeat signal is carried by these 4

component frequencies which seem to be

approximately integer multiple components of





1.08, as a fundamental frequency or first

harmonic. Therefore, we can conclude that period of

the heartbeat signal per minute is 1.08x60 = 64.8.

The numbers of pulses on Figure 4 (bottom) prove

this.

Figure 4: Recorded ECG signal and its FFT corresponding

signals which shows the periodicity of the ECG signal.

Having shown that the ECG signals obtained by

the employed sensor are indeed periodic (Figure 4),

we now first explain the testing scenarios in which

our data have been recorded. Then, we show the

effects of the modifications that we have applied to

the system of (Balakrishnan, 2013). Next, we give

the details of the comparison of our system against

the Balakrishnan et al.’s work (Balakrishnan, 2013).

4.1 Testing Scenarios

Five test subjects were asked to participate in testing

the systems from which 32 different videos were

recorded. These videos are recorded by a Logitech

webcam at a frame rate of 30 fps in different facial

expressions and head poses. These are the situations

in which the videos have been recorded in:

 Subjects look directly into the camera without

changing their facial expressions (This is the

same imaging condition as the system of

(Balakrishnan, 2013)).

 Subjects turn around their faces from left

(180

) to right (180

) and look at seven

different targets that are located at the same

distance from each other.

 Subjects show smiling/laughing expression.

 Subjects repeat a given sentence.

 Subjects show angry expression.

The duration of each video is around 60 seconds.

4.2 The Moving Average Filter and

DCT

Before obtaining the periodicity of the selected

source signal (Figure 1 block diagram), the only

difference between our system and the work of

(Balakrishnan, 2013) is that we have introduced a

moving average filter. This does not have much

effect when the face is standing still, and is facing

the camera. But, as soon as the subject is changing

his/her head pose and/or facial expressions are

changing, there will be so many occlusions in the

tracking of the feature points, that without using a

moving average filter the results will be erroneous.

Comparing Figure 5 (top) to Figure 5 (bottom)

shows that including this moving average filter

causes the employed PCA to pick a much smoother

signal as the strongest component compared to the

case where such a filter has not been included

(Balakrishnan, 2013). This will gives us better

results, for estimating the heartbeat rates, in the final

step of the system.

Figure 5: Comparing the estimated heartbeat rate signal

when the moving average filter is used (top) and when it is

not used (bottom).

Besides introducing the moving averge filter for

smoothing the estimated signal, in our system we

have used DCT to estimate the periodicity of the

estimated signal. The effect of this decision and

comparing it with the FFT of (Balakrishnan, 2013) is

shown in Figure 6. In this figure (top and middle

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

122

parts) a signal and its FFT representation are shown.

The maximum power of this FFT (3.603) gives a

heartbeat rate of 3.603x60 = 216.18, while the actual

heartbeat rate in this case is 60 bpm which can be

estimated much better using the first harmonic

(1.001x60=60.06). Therefore, the total spectral

power of the signal and then using the maximal

power and its first harmonic as have been used in

(Balakrishnan, 2013) does not always produce the

desired results. Instead, by using DCT in Figure 6

(bottom) it can be seen that a much better result will

be obtained, if the component number 20 is selected

as the component which carries the power of pulse

frequency. Feeding this value of this component in

the algorithm of section 3.4 results in an estimated

beat rate of 60.88 bpm, which is very close to the

actual value.

Figure 6: Extracting the beat rate of the signal (top) using

the algorithm of (Balakrishnan, 2013) (middle) and our

employed DCT (bottom).

4.3 Detailed Experiments

The proposed system has been compared against the

state-of-the-art work of (Balakrishnan, 2013) using

the testing data that was recorded in the previously

explained testing scenarios. The results of

comparing these systems (the proposed system and

the work of (Balakrishnan, 2013) against the ground

truth data obtained by the Shimmer ECG sensor for

the case which the testing subjects are looking

directly into the camera are shown in Table 1. In this

table, (a) is the subject number, (b) is the ground

truth data read by a Shimmer ECG device, (c) is the

heartbeat rate estimated by the system of

(Balakrishnan, 2013), (d) is the error of the method

of (Balakrishnan, 2013), (e) is the heartbeat rate

estimated by our proposed method, (f) is the error of

our proposed method. It can be seen that the error of

our system is generally better than that of

(Balakrishnan, 2013).

Table 1: The proposed system against system of

(Balakrishnan, 2013), please see the text for descriptions

of the headings.

a b c d e f

S1-1 61.71 63.06 1.35 62.1 0.39

S1-2 66.67 67.04 0.37 67.03 0.36

S2-1 60 216.83 156.8 61.88 1.88

S2-2 59 61.06 2.06 59.10 0.1

S2-3 54.00 53.03 0.97 54.11 0.11

S3-1 66.65 69.05 2.40 67.63 0.98

S4-1 84.06 86.06 2.00 83.90 0.16

S5-1 47.62 48.03 0.41 46.17 1.45

The size of the window employed for the moving

average filter in the previous experiment is set to

one. It means, no moving average is applied to the

data obtained from the previous test. Because, the

signal is already smooth. But, when it comes to the

case where facial expressions and/or head pose are

changing, the effect of the moving average becomes

more visible. Table 2 shows the results of the

proposed system against the work of (Balakrishnan,

2013) and the ground truth. The descriptions of the

headings (a)-(f) are the same as those for Table 1.

The size of the moving average window changes

between 40-80 samples, for different testing

scenarios. It can be seen from this table that the

proposed system is more robust than the work of

(Balakrishnan, 2013) in most of the cases, when the

facial expression and/or head pose are changing.

5 CONCLUSIONS

Motivated by the fact that in many applications like,

e.g., remote patient monitoring, there is not a

possibility for installing a device on the body of the

patients, this paper has proposed a contactless

heartbeat rate measurement using computer vision

techniques. The system finds some robust feature

points inside the facial areas of the users and tracks

them over time to generate some trajectories of the

feature points. These trajectories are then smoothed

by a moving average filter. Then, the irrelevant

frequencies are removed from the trajectories. All of

these refined trajectories are then fed to a PCA

algorithm to find the strongest independent

component. This component is assumed to be the

estimated heart beat signal. To find the periodicity of

this estimated signal a DCT-based algorithm has

been used. Experimental results on several video

0 1 2 3 4 5 6 7 8 9 10

-0.5

0.5

time(s)

Signal

0 1 2 3 4 5 6 7 8

5000

X: 3.603

Y: 4068

Frequency (Hz)

Power

0 50 100 150 200 250 300 35

-0.2

0.2

X = 20

Y = -0.105

Magnitue

X = 40

Y = 0.0956

ImprovedPulseDetectionfromHeadMotionsusingDCT

123

Table 2: Comparing the results of the proposed system against the state-of-the-art work of (Balakrishnan, 2013) when the

facial expressions and/or head pose are changing. Please see the text for the descriptions of the headings.

The test subject is smiling! The test subject is speaking!

a b c d e f c d e f

S1-1 66.67 49.051 17.61 58.22 8.45 81.28 14.61 67.31 0.64

S2-1 59 48.02 10.98 59.35 0.35 75.09 16.09 58.41 0.59

S2-2 54.00 56.04 2.04 54.99 0.99 79.05 25.05 52.66 1.34

S3-1 66.65 48.04 18.61 56.77 9.88 51.03 15.62 66.50 0.15

S4-1 84.06 148.09 64.03 61.73 22.33 61.04 23.02 63.78 20.28

S5-1 47.62 47.03 0.59 48.44 0.82 46.72 0.9 48.14 0.52

The test subject is angry! The head pose is changing!

a b c d e f c d e f

S1-1 66.67 50.05 16.62 59.23 7.44 49.03 17.64 60.47 6.2

S2-1 59 63.04 4.04 59.86 0.86 48.02 10.98 59.69 0.69

S2-2 54.00 49.04 4.96 63.96 9.96 50.04 3.96 53.04 0.96

S3-1 66.65 49.05 17.60 59.92 6.73 50.05 16.6 58.87 7.78

S4-1 84.06 63.03 21.03 47.76 36.3 146.10 62.04 57.87 26.19

S5-1 47.62 50.05 2.43 59.23 11.61 51.05 3.43 45.90 1.72

sequences show that the estimated heartbeat rates in

different facial expressions and head poses are very

close to the ground truth. Furthermore, it is shown

that the proposed system outperforms state-of-the-

art.

REFERENCES

Balakrishnan, G., Durand, F. and Guttag, J., 2013,

Detecting Pulse from Head Motions in Video.

Computer Vision and Pattern Recognition,

Proceedings CVPR '13, 2013 IEEE Computer Society

Conference on.

Bouguet, J., 2000, ‘Pyramidal implementation of the

Lucas Kanade feature tracker’, Intel Corporation,

Microprocessor Research Labs, Technical Report.

Hu, S., Zheng, J., Chouliaras, V. and Summers, R., 2008,

‘Feasibility of imaging photoplethysmography’, in

Proceedings of IEEE Conference on Biomedical

Engineering and Informatics, pp. 72–75.

Humphreys, K., Ward, T. and Markham, C., 2007,

‘Noncontact simultaneous dual wavelength

photoplethysmography: a further step toward

noncontact pulse oximetry’, Rev. Sci. Instrum., vol.

78, Issue 4, pp. 853-857.

Poh, M., McDuff, D. and Picard, R., 2010, ‘Non-contact,

automated cardiac pulse measurements using video

imaging and blind source separation’, Optic express,

Optics Express, Vol. 18, Issue 10, pp.10762-10774.

Shi, J., Tomasi, C., ‘Good features to track,’ 1994,

Computer Vision and Pattern Recognition,

Proceedings CVPR '94., 1994 IEEE Computer Society

Conference on, pp.593-600.

Takano, C. and Ohta, Y.,2007, ‘Heart rate measurement

based on a time-lapse image’, Med. Eng. Phys. Vol.

29, Issue 8, pp. 853-857.

Verkruysse, W., Svaasand, L. and Nelson, J., 2008,

‘Remote plethysmographic imaging using ambient

light’, Opt. Express, Vol. 16, Issue 26, pp. 21434-

21445.

Viola, P. and Jones, M.,2001, ‘Rapid object detection

using a boosted cascade of simple features’, Computer

Vision and Pattern Recognition, Proceedings CVPR

'01., 2001 IEEE Computer Society Conference on, pp.

511-518.

Wieringa, F., Mastik, F. and van der Steen, A., 2005,

‘Contactless multiple wavelength

photoplethysmographic imaging: a first step toward

“SpO2 camera” technology’, Ann. Biomed. Eng. Vol.

33, Issue 8, pp.1034-1041.

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

and Freeman, W., 2012 ‘Eulerian Video Magnification

for Revealing Subtle Changes in the World’, ACM

Transactions on Graphics, Proceedings SIGGRAPH

2012, Vol. 31, Issue 4, pp. 65-72.

VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications

124