A Smartphone Tool for Evaluating Cardiopulmonary Resuscitation

(CPR) Delivery

Gavin Corkery and Kenneth Dawson-Howe

School of Computer Science and Statistics, Trinity College, University of Dublin, Ireland

Keywords:

Video Processing, Machine Vision, Optical Flow, Cardiopulmonary Resuscitation.

Abstract:

This paper presents a prototype smartphone application to aid with the delivery of cardiopulmonary resuscita-

tion (CPR). The person giving CPR is viewed from the side and both compressions and breaths are identiﬁed

primarily using optical ﬂow. This allows the system to provide near real time feedback on the chest com-

pression rate (CCR) and on the timing of breaths (which affects the Chest Compression Fraction (CCF)). The

system is evaluated on over 25 minutes of video of 6 different participants delivering CPR to a test dummy. A

quantitative evaluation is presented which shows that the system recognised 99% of compressions and all of

the breaths (although two false positive breaths were classiﬁed). It computed the CCF to within 1%.

1 INTRODUCTION

The use of cardiopulmonary resuscitation (CPR) has

been shown to increase survival levels from 5.2% to

7.8% in the case of normal CPR and to 13.3% in

the case of Compression-only CPR (COCPR) (Bo-

brow et al., 2010). In the same study they found that

CPR or COCPR was only administered in around one

third (34.3%) of the out-of-hospital cardiac arrest in-

cidents. However it seems that this rate can be signi-

ﬁcantly increased (to around 60%) when proposed by

the emergency services telephone dispatcher (Dami

et al., 2010). In addition, it appears that if a deﬁbrilla-

tor (AED) is available the survival rate increases sig-

niﬁcantly. In a study by Iwani et al. (Iwami et al.,

2012) considering only cases where an public access

AED was used, the survival rate for normal CPR was

32.9% and for COCPR was 40.7%.

1.1 CPR Guidelines

The (American Heart Association (AHA)) re-

commendations for CPR are currently that chest

compressions (CC) should be applied at a rate (CCR)

of 100-120 per minute to a depth (CCD) of 5-6cm

allowing full recoil (decompression of the chest)

after each compression (AHA, 2015). The two hands

should be placed (one on top of the other) with the

heel of the lower hand in the centre of the chest on

the lower half of the breastbone (sternum), and com-

pressions should be done ﬁrmly and smoothly

pressing downward while keeping the arms straight.

For a layperson who is trained in giving rescue bre-

aths it is recommended that the chest compressions

should be alternated with rescue breaths in a ratio of

30 compressions to 2 breaths. Each breath should

be delivered over one second causing the chest to

rise, and the chest should fall again between breaths.

The period taken for the breaths reduces the chest

compression fraction (CCF - i.e. the portion of time

during which compressions are performed) which

should be as high as possible with a target of at least

60%.

1.2 Existing Systems for

Aiding/Evaluating CPR

The AHA guidelines state that it “may be reasonable

to use audiovisual feedback devices during CPR for

real-time optimization of CPR performance”, but no-

tes that “studies to date have not demonstrated a signi-

ﬁcant improvement in favorable neurologic outcome

or survival to hospital discharge with the use of CPR

feedback devices during actual cardiac arrest events”

(AHA, 2015). In recent years the range of techno-

logy available to provide CPR quality feedback has

signiﬁcantly expanded so there is some potential for

this feedback to improve positive outcomes. In addi-

tion such technologies provide methods for improving

CPR training.

Corkery, G. and Dawson-Howe, K.

A Smartphone Tool for Evaluating Cardiopulmonary Resuscitation (CPR) Delivery.

DOI: 10.5220/0007258904890496

In Proceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2019), pages 489-496

ISBN: 978-989-758-354-4

489

In lab/training scenarios, transmitter-receiver

pairs have been used for evaluating chest compression

depth (Kim et al., 2017; Kang et al., 2010) with high

precision. Also in a lab environments evaluation (of

CCR, CCD and CCF) has been done using RGB-D

sensors (Higashi et al., 2017; Loconsole et al., 2016)

and balance boards (Hayashi and Minazuki, 2017; Hi-

gashi et al., 2017; Ferreira et al., 2017) which can give

a measure of the force direction as well as the depth.

These preceding technologies are only appropriate for

use in lab/training environments but a number of new

technologies have potential in the ﬁeld.

Every AED has sensors built into it and these have

been used to measure both CCR and CCF (Torney

et al., 2016; Gonzlez-Otero et al., 2012). Accelero-

meters have been used independently (Yamamoto and

Ohmura, 2015; de Gauna et al., 2015), on smartwat-

ches (Ahn et al., 2016), or smartphones (Song et al.,

2015; Amemiya and Maeda, 2013) to evaluate CCD

and CCR. In addition smartphone cameras have been

used to evaluate CCR based on a view looking up-

wards at the person applying CPR (Meinich-Bache

et al., 2017; Frisch et al., 2014), where the smartp-

hone is lying ﬂat on the ground

There are hundreds of smartphone applications

which aid in the training and delivery of CPR (Ahn

et al., 2016) although most of these tools do not pro-

vide feedback regarding the quality of the CPR being

given. For CPR tools which do provide CPR quality

feedback, this is done as audio feedback, typically

in the form of a metronome (Amemiya and Maeda,

2013; Loconsole et al., 2016) or messages such as

“Push Faster”, “Good Speed”, “Push Slower”) (Tor-

ney et al., 2016) and as video feedback, typically a

colour indication of CCR quality (Ahn et al., 2017;

Amemiya and Maeda, 2013).

Only one piece of research (that we could locate)

considers the posture of the person performing CPR

in terms of keeping the arms straight (Higashi et al.,

2017). No research could be located which consider

the position of the hands on the chest of the patient.

In addition Panicker et al. presented work identi-

fying CPR scenes in medical. simulation videos (Pa-

nicker et al., 2015)

1.3 System Concept

The research presented in this paper describes the cre-

ation and evaluation of the ﬁrst version of a smartp-

hone application intended to aid in the delivery of

CPR. The concept of the system is that it should as-

sess the CCR (and the breathing phases) based on the

analysis of a video of a person giving CPR taken on

a smartphone. This evaluation must happen as close

Figure 1: Dense optical ﬂow showing downward movement

(left) and upward movement (right).

to real-time as possible and provide feedback on the

CCR so that the person can increase or decrease the

compression rate appropriately.

2 SYSTEM

To recognise both the compressions and the breaths

we rely on dense optical ﬂow (as implemented in

OpenCV); See Figures 1 and 2.

In order to eliminate some noise, the amount of

movement from frame to frame must meet a certain

threshold before it is considered. This reduces the im-

pact of very small movements on the algorithm. Exa-

mining video footage showed a great amount of mo-

vement in the compressions, so a small threshold does

not limit the recognition of this movement. The thres-

hold was chosen to be quite small, requiring at least

0.5 pixels of movement between frames. This recog-

nises the moving region of interest while reducing a

lot of noise.

In order to reduce noise from the movement of

background objects, a model of the moving and non-

moving regions of the scene is maintained. As the

CPR session progresses, the regions with a great

amount of historical movement are given a greater

weighting over those that have been historically sta-

tic, meaning that background movement will cause

fewer false positives. The weighting is updated at

each frame for each pixel by the amount of movement

from the previous frame, multiplied by a small lear-

ning rate (experimentally chosen to be 0.005). After

several compressions, there will be a clear model of

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

490

the moving regions of the image, which include the

head, chest and arms of the person giving CPR. The

weighting is not updated during the breathing phase,

as the movement model is focused solely on the mo-

vement during compressions.

2.1 Recognising Compressions

Having a weighted motion ﬁeld for each frame

forms the basis for recognising chest compressions.

Through analysis of test videos, it is clear that a chest

compression can be described by two phases: a strong

upward motion phase and a strong downward motion

phase (See Figure 1). A compression is deemed to be

found when strong upward movement occurs after a

strong downward movement. The upward movement

must be within a small window of time after the do-

wnward movement.

Compressions are characterised by both strong up-

ward and downward movement. To account for noisy

scenes, the dominant movement must also be several

times (the ratio chosen was 3) greater than the mo-

vement in the other direction. This ratio is useful in

a scene with a lot of background movement. There is

also a threshold for the amount of the scene that must

be moving. This was chosen to be 10% to ensure that

the movement takes up at least a small portion of the

scene.

In order to reduce false positives, the timing of the

upward movement in relation to the downward mo-

vement is also measured. For a compression to be

recognised, the upward movement must take place no

longer than 1 second after the downward movement.

This value was chosen to reduce false positives, ho-

wever it may cause difﬁculties with recognising ex-

tremely slow compressions.

2.2 Recognising Breaths

The recognition of breaths uses a similar technique to

the recognition of compressions. The recognition of a

breath is split into two phases: the downward lateral

movement and the upward movement at the end of the

breath (See Figure 2).

The lateral movement is recognised by using a

threshold based on the mean lateral movement du-

ring compressions. The total movement to the left

or right must be 5 times greater than the mean mo-

vement in that direction. The lateral movement must

also be at least a third of the downward movement.

This movement must take place in at least 3 conse-

cutive frames to be counted as the start of a breath.

This is to ensure that the movement is actually taking

place. The upward movement at the end of the bre-

Figure 2: Rescuer movement while starting breaths (left),

during breaths (middle) and moving back into compressions

(right).

Figure 3: Simpliﬁed ﬂow of vision algorithm upon recei-

ving a new frame.

ath is recognised in the same way that the end of a

compression is recognised. This also acts as a mecha-

nism for recovering from error if a breath was falsely

identiﬁed.

The overall algorithm for processing individual

frames is illustrated in Figure 3.

2.3 Calculating CCR and CCF

In order to provide useful feedback, the application

calculates the CCR while performing compressions,

and the CCF when the CPR session has concluded.

The CCR is calculated using a weighted learning

approach. For each compression, the calculated rate

is computed as 36000 (the number of milliseconds in

a minute) divided by the number of milliseconds be-

tween compressions. This means that if there is half

a second (500 milliseconds) between two compressi-

ons, the rate is 120 compressions per minute. The

A Smartphone Tool for Evaluating Cardiopulmonary Resuscitation (CPR) Delivery

491

”time” of a compression is deﬁned as the time of the

ﬁrst strong upward movement after downward mo-

vement.

For the ﬁrst 5 compressions, the impact of each

new compression is 50% of the overall rate. This al-

lows for a quick establishment of the rate at the start

of compressions. After the ﬁrst 5 compressions, each

new compression makes up 15% of the overall rate.

The reason for this lower rate is to minimise the user’s

oscillation between slow and fast compressions due to

negative feedback causing vast overcorrections.

The CCF is calculated using the following formula

CCF =

(Totaltime − InterruptedTime)

Totaltime

In this case, the interrupted time refers to the sum

of the periods of time which were at least 2 seconds

long and did not contain any compressions. This will

include time taken for breaths.

3 MOBILE APPLICATION

The implementation of the aforementioned techni-

ques is realised through a smartphone application.

The landing page of the application provides the user

with a view of the current scene through either the

front or rear camera, depending on user selection. In

training/practice situations, the smartphone should be

placed at a close distance from the CPR dummy, with

the CPR dummy lying between the smartphone and

the rescuer. Since different smartphone cameras will

have different ﬁelds of view, a precise optimal dis-

tance from the phone cannot be recommended. Ide-

ally, the rescuer and dummy should both be clearly

visible and ﬁll a majority of the camera view.

3.1 Realising Acceptable Frame Rate

The most important aspect to consider in an applica-

tion processing live video is the resulting frame rate

after processing each frame, which can be quite ex-

pensive. Some of the videos were of a very high re-

solution (1920x1080) and this resulted in a frame rate

of approximately 1 frame per second (fps). To handle

this, each image frame is resized to a 216x216 pixels

frame, which is over 25 times smaller. This gives

a massive performance boost without degradation to

the computer vision algorithm. The resulting perfor-

mance increase affords a frame rate of approximately

15fps, which we found to be sufﬁcient to provide high

quality feedback to the user. In addition, for speed

purposes, the vision algorithm does not analyse every

pixel, but samples at a constant rate. This is due to

the observation that the moving area will be a large,

mainly contiguous area. Analysing every pixel is the-

refore unnecessary so we analyse every 16th pixel in

each row or column is analysed, leading to

of mo-

vement to analyse. This spacing works well for all

tested distances, but may be too wide if the rescuer is

extremely far from the camera. The assumption, ho-

wever, is that the rescuer will never be so far away

from the camera for this to be an issue.

3.2 Methods of Feedback

As the person giving CPR would be focused on the

task, it is important that feedback is clear and does

not require careful analysis of the phone screen. The

visual feedback is as obvious and legible as possible,

while audio feedback is used to provide evaluation

without having to focus on the smartphone screen.

The information needed to provide metrics are pro-

cessed after the previous camera frame has been pro-

cessed, ensuring that feedback is always as up-to-date

as possible.

3.2.1 Visual Feedback

The primary measurable metric of CPR by the appli-

cation is the chest compression rate. When perfor-

ming CPR compressions, this is the aspect that should

be the main focus of the administrator’s attention. The

region at the top of the screen is used to provide an

easily-readable indicator of their current rate. The rate

is shown as a large number, which is colour-coded to

align with the recommended compression rate (See

Figure 4). Rates between 100 and 120 compressi-

ons per minute are optimal, and are thus displayed

with a green colour. Rates within 10 compressions

per minute of being optimal are shown with orange,

and rates further outside the optimal range are shown

in red. The displayed rate is updated after every com-

pression. This is done to ensure that any large rate

changes can be rectiﬁed as soon as possible.

Visual feedback is also provided to tell the user to

start breaths or to resume compressions (See Figure

5).

3.2.2 Audio Feedback

In cases where the person giving CPR may not be able

to focus their attention on the phone screen, it is pre-

ferable to have another way of delivering feedback in

a timely manner. This is achieved by using a speech

synthesiser giving feedback with a clear voice.

The audio feedback is related to the same pertinent

information dealt with by the visual feedback. Du-

ring the compression phase, the application will tell

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

492

Figure 4: Screenshots from the smartphone app showing

colour coded rates: optimal compression rate (left), sub-

optimal rate (middle), poor compression rate (right).

Figure 5: Screenshots from the smartphone app showing vi-

sual cues to perform breathing (left), that breathing is hap-

pening (middle), and that breaths have gone for too long and

compressions need to resume (right).

the user to keep going at their current rate if they are

achieving an optimal rate. If they are slightly out of

the optimal range they will be told to speed up slightly

or slow down slightly, and if they are very far out of

the range they will be told to speed up or slow down.

The frequency at which the user receives feedback on

their compression rate can be slow, medium or fast

(feedback every 6, 4 or 2 seconds respectively).

Feedback is also provided for other important as-

pects. If the user chooses to perform the 30:2 cycles

detailed above, they can choose to be notiﬁed to start

artiﬁcial ventilation every 30 compressions. Along

with this, they will be notiﬁed if their breaths have ta-

ken too long, and told to continue compressions. The

audio aspect of this is important, as the user would not

be focused on the phone screen if they are performing

artiﬁcial ventilation.

3.3 Viewing Summaries

At the conclusion of the CPR session, the administra-

tor is presented with a more in-depth analysis on the

summary screen (See Figure 6).

The goal of the summary screen is to provide furt-

her insight into the user’s habits that may not be appa-

rent from their CPR session. Examples of this could

Figure 6: Screenshot from the smartphone app showing

summary after CPR compressions.

be that the user starts their compressions too slowly

during each compression cycle, or that they are per-

forming too few or too many compressions per cycle.

Similarly to the visual feedback, the compression rate

per cycle is colour-coded in the same rate ranges. The

compression rate over time is graphed. If the user has

improved their CPR quality signiﬁcantly over time,

this should be identiﬁable through analysis of current

and past CPR sessions.

4 EVALUATION

The performance of the computer vision algorithm

was measured on 11 pre-recorded videos of CPR

being performed by ourselves and also by volun-

teers from the Dalkey First Responders Group. This

footage consists of over 25 minutes of CPR with bre-

athes being delivered to test dummies (8 videos with

resolution 768x576 pixels 25 fps, and 3 videos with

resolution 1920x1080 at 30 fps). To evaluate the per-

formance of our system we created ground truth for

this footage in terms of the frame numbers of the lo-

west point of each compression, and the presence of

A Smartphone Tool for Evaluating Cardiopulmonary Resuscitation (CPR) Delivery

493

breaths between sets of compressions.

4.1 Evaluation of Individual

Compressions and Breaths

When performing the overall calculations for the test

videos, the following formulae for accuracy, recall

and precision are used (where TP is a True Positive:

A compression or breath which occurred and was re-

cognised by the system; FP is a False Positive: A

compression or breath which was recognised by the

system, which did not actually occur; FN is a False

Negative: A compression or breath which did occur

but was not recognised by the system):

Accuracy =

T P

T P + FP + FN

Accuracy deﬁnes the proportion of times that the clas-

siﬁcation is correct.

Recall =

T P

T P + FN

Recall deﬁnes the proportion of breaths and compres-

sions which are correctly identiﬁed.

Precision =

T P

T P + FP

Precision deﬁnes the proportion of times the classiﬁ-

cation is correct when it says a breath or compression

has been identiﬁed.

4.2 Results

4.2.1 Compressions

The results for recognising compressions are very

good. The vision algorithm has an accuracy, preci-

sion and recall of 0.99. The false positives mainly

occur at the beginning and ending of the CPR ses-

sion, where the rescuer is kneeling down and posi-

tioning their hands, or standing up and exiting the

frame. This occurs due to the upward or downward

movement within the scene, which is falsely attribu-

ted to compressions. This will not affect the live rate

hugely, but it will affect the compression statistics on

the summary page. False negatives (missed compres-

sions) occur in two test videos. One reason for this is

a large amount of lateral movement is mistakenly as-

sumed to be the start of artiﬁcial ventilation, and sub-

sequent compressions are missed before the algorithm

self-corrects. The other reason is that compressions in

one video did not meet the movement threshold.

The results for recognising breaths are also very

good. Most notable is the fact that there are no missed

breaths. If missed breaths were to occur it is possible

that the overall rate would become inaccurate due to a

large gap between compressions without recognising

breaths in between them.

4.2.2 Evaluation of Chest Compression Rate and

Chest Compression Fraction

In all cases, the chest compression rate is sampled

every 2 seconds from the beginning of compressions

in the ground truth. This means that the results will in-

dicate the rate at the exact same points in time to give

a fair comparison. The chosen interval is 2 seconds

as this is the quickest option available for audio feed-

back on compression rates. The metrics used were

both the mean compression rate calculated throughout

each test video, but also the mean difference between

each calculated rate and the percentage of rates within

3 compressions per minute of each other. It is not only

important for the mean rates to be close, but for the

majority of rates at any point in time to be close to

one another. The results were very good (See Table

2), with the mean rates all within 1 compression per

minute of each other, and over 95% of rates in each

video being within 3 compressions per minute of the

real value. This suggests the calculated rate is good

enough to be considered an accurate assessment.

Discrepancies in the CCF between the vision algo-

rithm and the ground truth are mainly due to false po-

sitives at the start and end of the CPR session. These

false positives will falsely indicate that the CPR ses-

sion was longer than it actually was, giving a greater

overall fraction. The results for the CCF are encoura-

ging, with most results being within 1% of the true

result. In all cases, the computed CCF gives a rea-

sonable idea of whether the compressions are being

done for a suitable portion of time.

4.3 Comment on Results

The system gives extremely good results. Given that

there are very few false positives or false negatives

for compressions and breaths, it follows that the cal-

culated CCR and CCF are also very close to each ot-

her. The vision algorithm is not only able to recognise

breaths and compressions, but also of giving good re-

sults of the current rate at most points in time during

the CPR administration.

5 CONCLUSIONS

The prototype system presented here successfully

monitors CCR and breaths for people giving CPR. It

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

494

Table 1: True positives (TP), false positives (FP) and false negatives (FN) of compressions and breaths for each test video,

along with their totals. In addition the overall accuracy, precision and recall for compressions and breaths are presented.

Compressions Breaths

Video # TP FP FN TP FP FN

1 144 0 0 4 0 0

2.1 127 0 1 0 1 0

2.2 174 1 5 0 1 0

3 169 1 0 5 0 0

4 186 1 0 6 0 0

5.1 211 3 0 7 0 0

5.2 211 0 0 7 0 0

6.1 175 1 0 5 0 0

6.2 200 4 0 7 0 0

6.3 170 1 0 5 0 0

6.4 201 2 0 7 0 0

Totals 1974 14 6 53 2 0

Overall Accuracy Precision Recall Accuracy Precision Recall

0.99 0.99 0.99 0.96 0.96 1

Table 2: Mean CCR calculated from the ground truth and vision algorithm for each video, along with the proportion of rates

which were with 3 compressions per minute. In addition the CCF from the ground truth and from the system are presented.

Ground Truth Computed Proportion within Ground Truth Computed

Video # Mean CCR Mean CCR 3 compressions/min CCF CCF

1 108.00 108.15 1.00 0.773 0.773

2.1 103.80 103.90 0.95 0.958 0958

2.2 73.95 73.60 0.93 1.000 1.000

3 123.27 122.98 0.99 0.558 0.555

4 125.55 125.74 1.00 0.700 0.700

5.1 115.10 114.28 0.93 0.696 0.701

5.2 114.98 114.60 0.96 0.731 0.732

6.1 92.50 92.50 1.00 0.800 0.796

6.2 92.17 91.60 0.95 0.795 0.807

6.3 92.84 92.69 1.00 0.800 0.791

6.4 92.22 91.59 0.96 0.771 0.807

also provides an accurate value of the CCF. As this is

the ﬁrst stage in a larger development it has not yet

been tested for robustness with respect to the position

of the camera relative to the person giving CPR.

The concept clearly has potential for use in

training environments, where volunteers typically

practice CPR every time they meet but rarely with

precise feedback on CCR or CCF. In the longer term

this technique might have potential for use in the ﬁeld

although a number of other issues need to be addres-

sed (such as CCD, robustness, camera location requi-

rements, etc.).

REFERENCES

AHA (2015). 2015 American Heart Association (AHA)

Guidelines Update for CPR and Emergency Cardio-

vascular Care (ECC). American Heart Association.

Ahn, C., Cho, Y., Oh, J., Song, Y., Lim, T. H., Kang, H.,

and Lee, J. (2016). Evaluation of smartphone appli-

cations for cardiopulmonary resuscitation training in

south korea. BioMed Research International, 2016:1–

Ahn, C., Lee, J., Oh, J., Song, Y., Chee, Y., Lim, T. H.,

Kang, H., and Shin, H. (2017). Effectiveness of feed-

back with a smartwatch for high-quality chest com-

pressions during adult cardiac arrest: A randomized

controlled simulation study. PloS one, 12:e0169046.

Amemiya, T. and Maeda, T. (2013). Poster: Depth and

rate estimation for chest compression cpr with smartp-

hone. In 2013 IEEE Symposium on 3D User Interfaces

(3DUI), pages 125–126.

A Smartphone Tool for Evaluating Cardiopulmonary Resuscitation (CPR) Delivery

495

Bobrow, B. J., W Spaite, D. W., Berg, R. A., Stolz, U.,

Sanders, A. B., Kern, K., Vadeboncoeur, T. F., Clark,

L. L., Gallagher, J., Stapczynski, J. S., Lovecchio,

F., Mullins, T. J., Humble, W. O., and Ewy, G. A.

(2010). Chest compression-only cpr by lay rescu-

ers and survival from out-of-hospital cardiac arrest.

JAMA : the journal of the American Medical Asso-

ciation, 304:1447–54.

Dami, F., Carron, P.-N., Praz, L., Fuchs, V., and Yersin,

B. (2010). Why bystanders decline telephone cardiac

resuscitation advice. Academic Emergency Medicine,

17(9):1012–1015.

de Gauna, S. R., Gonzlez-Otero, D. M., Ruiz, J., Chicote,

B., and Vidales, N. (2015). Alternatives to estimate

the compression depth from the acceleration signal

during cardiopulmonary resuscitation. In 2015 Com-

puting in Cardiology Conference (CinC), pages 561–

564.

Ferreira, L. R., Kapps, G. W., de Oliveira, J. C., and Shir-

mohammadi, S. (2017). An instrument for measuring

force vector and frequency of cpr compressions. In

2017 IEEE International Instrumentation and Measu-

rement Technology Conference (I2MTC), pages 1–5.

Frisch, A., Das, S., Reynolds, J., De La Torre, F., Hodgins,

J., and Carlson, J. (2014). Analysis of smartphone vi-

deo footage classiﬁes chest compression rate during

simulated cpr. American Journal of Emergency Medi-

cine, 32(9):1136–1138.

Gonzlez-Otero, D., Ruiz de Gauna, S., Ruiz, J., Ayala,

U., and Alonso, E. (2012). Automatic detection of

chest compression pauses using the transthoracic im-

pedance signal. In 2012 Computing in Cardiology,

pages 21–24.

Hayashi, H. and Minazuki, A. (2017). Development

of a cardiopulmonary resuscitation feedback lear-

ning device using maltisensor. In 2017 14th

International Conference on Electrical Engineer-

ing/Electronics, Computer, Telecommunications and

Information Technology (ECTI-CON), pages 473–

476.

Higashi, E., Fukagawa, K., Kasimura, R., Kanamori, Y.,

Minazuki, A., and Hayashi, H. (2017). Development

and evaluation of a corrective feedback system using

augmented reality for the high-quality cardiopulmo-

nary resuscitation training.

Iwami, T., Kitamura, T., Kawamura, T., Mitamura, H., Na-

gao, K., Takayama, M., Seino, Y., Tanaka, H., No-

nogi, H., Yonemoto, N., and Kimura, T. (2012). Chest

compression–only cardiopulmonary resuscitation for

out-of-hospital cardiac arrest with public-access deﬁ-

brillation. Circulation, 126(24):2844–2851.

Kang, M. K., Oh, J. H., and Kim, T. (2010). Chest compres-

sion depth measurement using iruwb for cpr (cardio-

pulmonary resuscitation). In 2010 IEEE Asia Paciﬁc

Conference on Circuits and Systems, pages 496–499.

Kim, Y., Yu, B. G., Oh, J. H., and Kim, T. W. (2017). Novel

chest compression depth measurement sensor using ir-

uwb for improving quality of cardiopulmonary resus-

citation. IEEE Sensors Journal, 17(10):3174–3183.

Loconsole, C., Frisoli, A., Semeraro, F., Stroppa, F., Ma-

stronicola, N., Filippeschi, A., and Marchetti, L.

(2016). Relive: A markerless assistant for cpr trai-

ning. IEEE Transactions on Human-Machine Sys-

tems, 46(5):755–760.

Meinich-Bache, Ø., Engan, K., Eftestøl, T., and Austvoll,

I. (2017). Detecting chest compression depth using

a smartphone camera and motion segmentation. In

Lecture Notes in Computer Science, volume 10270,

pages 53–64.

Panicker, M. S. A., Frigui, H., and Calhoun, A. W. (2015).

Identiﬁcation of cardio-pulmonary resuscitation (cpr)

scenes in medical simulation videos using spatio-

temporal gradient orientations. In 2015 International

Conference on Image Processing Theory, Tools and

Applications (IPTA), pages 365–369.

Song, Y., Oh, J., and Chee, Y. (2015). A new chest compres-

sion depth feedback algorithm for high-quality cpr

based on smartphone. Telemedicine journal and e-

health, 21:36–41.

Torney, H., O’Hare, P., Davis, L., Delafont, B., Bond,

R., McReynolds, H., McLister, A., McCartney, B.,

Maio, R. D., and McEneaney, D. (2016). A usability

study of a critical man machine interface: Can lay-

person responders perform optimal compression rates

when using a public access deﬁbrillator with automa-

ted real-time feedback during cardiopulmonary resus-

citation? IEEE Transactions on Human-Machine Sy-

stems, 46(5):749–754.

Yamamoto, K. and Ohmura, R. (2015). Towards on-site in-

struction of cardio pulmonary resuscitation with wea-

rable sensors. In 2015 IEEE/SICE International Sym-

posium on System Integration (SII), pages 387–392.

VISAPP 2019 - 14th International Conference on Computer Vision Theory and Applications

496