Measuring Respiration and Heart Rate using Two Acceleration Sensors

on a Fully Embedded Platform

Johan Vertens

, Fabian Fischer

, Christian Heyde

, Fabian Hoeﬂinger

, Rui Zhang

Leonhard Reindl

and Albert Gollhofer

EMP, Albert-Ludwigs-University, Georges-Koehler-Allee 106, 79110 Freiburg, Germany

Institut fuer Sport und Sportwissenschaft, Albert-Ludwigs-University, Schwarzwaldstraße 175, 79117 Freiburg, Germany

Keywords:

Respiration Rate, Heart Rate, Acceleration Sensor, Optimization, Bluetooth, Wearable Device.

Abstract:

We present a novel system which measures the respiration rate using two three-axis accelerometers strapped

to the chest and the back of a person. Respiration and heart rate are used as a measurement to determine

an athlete’s ﬁtness level during the exercise phase. Common respiration rate measurement methods require

devices which are mostly stationary or at least clunky and uncomfortable to wear for an extended period of

time.

By using techniques such as differential measurement, pre-measurement optimization, adaptive ﬁltering and

peak detection we are able to obtain respiration rate even when the athlete is running fast. Our system is

low-cost, small, and, by using a digital signal processor, fully capable of processing the data in an online way.

Results are compared with a reference spirometer and have shown a very low normalized root mean square

error (NRMSE) down to 1.42 %. Additionally, the heart rate of an athlete can be measured with the same

sensor setup and similar methods.

1 INTRODUCTION

Ambulant and user friendly methods to derive an indi-

vidual’s physiological demand during rest and phys-

ical exercise are currently based on measurements of

heart rate. However, heart rate mainly provides in-

formation about the cardiac loading rather than infor-

mation about the metabolic strain - as for instance by

certain stress thresholds derived from changes in the

blood lactate concentration or the respiratory gas ex-

change (Wasserman, 2012). Thus, the accurate as-

sessment of an individual’s aerobic ﬁtness level as

well as the precise determination of an individual’s

intensity level for aerobic exercise are feasible only

using a laboratory setting. The examination of the res-

piratory gas exchange usually requires clunky face-

masks or mouthpieces which are uncomfortable to

wear for an extended period of time to derive the

needed measurements at the oriﬁce.

Interestingly, there are some promising attempts

demonstrating that simply the observation of changes

in the respiration rate may facilitate stress threshold

detection (Onorati et al., 2012; Carey et al., 2005).

Deriving respiration rate during rest and exercise does

not necessarily need the measurement at the oriﬁce

but can alternatively be derived from thoracic surface

displacements (Heyde et al., 2014). Furthermore, the

measurement of respiration rate solely needs the pre-

cise detection of the onsets of inspiration and expi-

ration within a sinusoidal like data pattern. There-

fore, it is hypothesized that common sensors that

are surrounding the whole thorax (i.e. inductance

plethysmographs) might be replaceable by smaller ac-

celerometers which are able to register breathing in-

duced thorax extension. In this pilot study, respiration

rate values derived by means of an dual accelerome-

ter system mounted at the chest and a gold standard

ﬂowmeter placed at the oriﬁce were compared. Addi-

tionally, the system was used to simultaneously derive

heart rate values in a similar way.

2 STATE OF THE ART

Currently, the respiration information can be mea-

sured by using many different sensor techniques. Our

focus is to use inertial sensors for detecting the respi-

ration rate. Inertial sensors are already used for dif-

ferent applications (Simon et al., 2015; Hoeﬂinger

Vertens, J., Fischer, F., Heyde, C., Hoeﬂinger, F., Zhang, R., Reindl, L. and Gollhofer, A..

Measuring Respiration and Heart Rate using Two Acceleration Sensors on a Fully Embedded Platform.

In Proceedings of the 3rd International Congress on Sport Sciences Research and Technology Support (icSPORTS 2015), pages 15-23

ISBN: 978-989-758-159-5

et al., 2012a; Hoeﬂinger et al., 2011). In applica-

tions for human tracking they are already integrated

into shoes or clothes (Hoeﬂinger et al., 2012b; Zhang

et al., 2013) for detecting the body movement and

measuring the path. Inertial sensors have been in-

creasingly used in recent years to derive respiration

rate. Accelerometers worn on the torso are capable

of measuring inclination and angular changes during

respiration. Afterwards, the respiration rate can be

estimated using digital signal processing. Liu et al.

present a method using adaptive band-pass ﬁlter and

principal component analysis (PCA) to derive the res-

piratory rate from acceleration data (Liu et al., 2011).

The method was capable of offering dynamic respi-

ration rate estimation during various body activities

such as sitting, walking, running, and sleeping. Tewel

presents a new device for detection of apnoea, con-

sisting of a three-axis MEMS accelerometer with dig-

ital output, microprocessor and some alarm instru-

ments (Tewel, 2010). A wireless portable monitoring

system to measure a user’s respiratory airﬂow, blood

oxygen saturation, and body posture is proposed in

(Cao et al., 2012). The monitoring system consists of

two sensor nodes including a hot-ﬁlm ﬂow sensor, tri-

axis accelerometer and oximeter. Phan et al. used an

accelerometer to measure cardio-respiratory activity

(Phan et al., 2008). The acquisition is realized in dif-

ferent modes: normal, apnoea, deep breathing or af-

ter exhaustion and also in different postures: vertical

(sitting, standing) or horizontal (lying down). Yoon et

al. suggest a method to improve the fusion of an ac-

celerometer and a gyroscope by using a Kalman ﬁlter

to produce a higher quality respiration signal (Yoon

et al., 2014). The authors acclaim that the acceleration

signal due to the movement can be easily removed

because the frequency of movement acceleration is

much higher than the frequency of respiration. How-

ever, it was found not true during our ﬁrst attempt. Jin

et al. proposed and analyzed three different methods

to extract a single respiratory signal from the tri-axial

data (Jin et al., 2009). The system is evaluated us-

ing simulated data from the most common postures,

such as lying and sitting, as well as real data collected

from ﬁve subjects. Bates et al. use a movement de-

tection method to classify periods in which the patient

is static and breathing signals can be observed accu-

rately (Bates et al., 2010).

3 METHODS AND MATERIALS

3.1 Measurement Principle

With every breath a human takes his chest expands.

We use two three-axis accelerometers centered at the

front and the back of the torso of a person to measure

the acceleration and inclination caused by the expan-

sion of the chest. Both sensors are strapped to the

torso with a ﬂexible belt. Figure 1 shows the setup.

Figure 1: Setup of the two sensors (red). Translational and

rotational directions are shown as arrows.

By using two sensors we can apply the method of

differential measurement of the acceleration a, i.e.

a = a

front

−a

back

. If a perfect alignment of the coordi-

nate systems of both sensors is achieved, this method

eliminates acceleration measurements introduced by

translational movement which do not belong to the

respiration.

After transformation and ﬁltering of the data we

detect the peaks of every amplitude and calculate the

respiration and heart rate from the time difference be-

tween two adjacent peaks.

3.2 Hardware

For measuring and processing the acceleration data

we designed two printed circuit boards which contain

all the relevant digital blocks. The front pcb, which

is mounted on the chest, contains the primary com-

ponents as a STM32F4 microprocessor with an on-

board DSP, a low energy bluetooth chip, a voltage reg-

ulator, ﬂash memory and a LIS3DSH accelerometer.

The back pcb holds the second LIS3DSH accelerom-

eter and a port for connecting the two modules with

each other. We have selected the sensors due to their

very high sensitivities of 0.06 mg per bit in the mea-

surement range ±2.0 g with 16-bit data output. All

communication between the digital parts is realized

through SPI interface. Energy is delivered to the sys-

tem via a small lithium battery. Communication to a

PC or to a mobile device can be established via a se-

rial or the bluetooth 4.0 connection. Since we have

a very powerful setup we are able to perform all the

signal processing onboard.

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

The size of the front pcb is 4.5 x 3.5 cm and con-

tains also the mounting tabs for the chest strap while

the back pcb is 2.3 x 2.0 cm. The hardware modules

are shown in ﬁgure 2 and as a block diagram in ﬁgure

Figure 2: Hardware module for chest-mounting (left) and

for back-mounting (right).

Cortex F4 with

DSP

Flash memory

16MB

3 Axis

Accelerometer

LIS3DSH

(Front)

Bluetooth Low

Energy Chip

Bluenrg

Matching

Network

3 Axis

Accelerometer

LIS3DSH

(Back)

Antenna

3.3V Voltage

Regulator

SPI

SPI SPI

Figure 3: Block diagram of the hardware setup with the

STM32F4 microprocessor, the two LIS3DSH accelerome-

ters, the ﬂash memory, the voltage regulator and the blue-

tooth chip with matching network and antenna. All commu-

nication between the digital parts is realized through SPI.

3.3 Calibration and Optimization

Slippage or unprecise mounting of the chest strap

causes misalignment of the coordinate systems be-

tween the front sensor and the back sensor. Since we

need perfect or at least very good alignment to gen-

erate the differential signal, we have developed a cal-

ibration routine for this purpose. In the case of mis-

alignment, the differential measurement is not reliable

any more. This is caused by accelerations which ap-

pear on different axes in the measurements of the two

accelerometers.

At ﬁrst the sensitivities and offsets of the sensors

were calibrated following the least squares approach,

whis is described in (STMicroelectronics, 2014). To

align the two sensors we rotate the back sensor virtu-

ally until the coordinate systems are overlapping. For

this procedure the rotation is described as rotation ma-

trices for the x-, y- and z-axis with the corresponding

angles Φ,Θ and Ψ. The coordinate system is visual-

ized in ﬁgure 4.

Figure 4: Coordinate system of an accelerometer.

(Φ) =





1 0 0

0 cos(Φ) −sin(Φ)

0 sin(Φ) cos(Φ)





(1)

(Θ) =





cos(Θ) 0 sin(Θ)

0 1 0

−sin(Θ) 0 cos(Θ)





(2)

(Ψ) =





cos(Ψ) −sin(Ψ) 0

sin(Ψ) cos(Ψ) 0

0 0 1





(3)

The total rotation matrix for all three axis can then be

formed by multiplying all matrices together:

A = R

(Φ

∗

) · R

(Θ

∗

) · R

(Ψ

∗

) (4)

Virtual rotation of the back sensor can be done by

rotating the measurements x, y and z of the back ac-

celerometer S2 with the rotation matrix A to the trans-

formed measurements ´x, ´y and ´z:









= A ·









(5)

If the rotation matrix is applied, the measurements of

the front sensor S1 and the transformed measurements

of the back sensor S2 are equal and the coordinate

systems are overlapping:

















(6)

Measuring Respiration and Heart Rate using Two Acceleration Sensors on a Fully Embedded Platform

For a rough guess of the rotation matrix A we use the

angular information which we directly get from the

accelerometer measurements, namely the earth grav-

ity vector. While standing still and holding the breath,

there is no other acceleration affecting the measure-

ment than the acceleration from gravity. In this situa-

tion the rotation angles Φ and Θ of the x- and y-axis

for each sensor S1 and S2 can be determined with the

following equations and the corresponding accelera-

tion values a

and a

S1,S2

= tan

−1

+ a

(7)

S1,S2

= tan

−1





+ a





(8)

After calculating these angular guesses one can con-

struct the angular differences Φ

and Θ

between the

front and the back sensor by simple subtracting:







− Φ

− Θ



(9)

By using Φ

and Θ

in the corresponding equations

1 and 2 one can calculate the estimated rotation ma-

trices, which transform the back sensor’s coordinate

system in order to fulﬁll equation 6. Since the wearer

of the belt will probably do some motion in the cal-

ibrating phase which might distort the angle calcula-

tion and since the previously described method can

only estimate the angles of the x- and y-axis, we fur-

ther implemented a more advanced calibration, which

takes the estimated angles, calculated from the earth

gravity vector, as an initial guess.

3.3.1 Optimized Calculation of Rotation

In the more advanced approach for calculating the

misalignment of the coordinate systems we introduce

the cost function C(Φ, Θ, Ψ) (equation 10) which cal-

culates how good the ﬁt between the measurements of

the front sensor S1 and the back sensor S2 is.

C(Φ, Θ, Ψ) =

∑

(S1)

− R

(Φ)x

(S2)

+|y

(S1)

− R

(Θ)y

(S2)

| (10)

+|z

(S1)

− R

(Ψ)z

(S2)

After collecting data for some seconds the cost func-

tion can be optimized over the rotation angles Φ,Θ

and Ψ (equation 11), in order to get the optimized an-

gles Φ

∗

,Θ

∗

and Ψ

∗

. For getting useful data one must

still hold the breath in the collecting phase since we

just want information which is about static position-

ing on the body. Dynamic movement along an axis

of the body while collecting measurement data helps

the optimizer to calculate a good result in contrast to

the previous estimation approach, presented in chap-

ter 3.3, where no movement should occur.





∗





= arg min

Φ,Θ,Ψ

C(Φ, Θ, Ψ) (11)

Since the processing speed of a microcontroller is

limited we used the stable and easy to implement lo-

cal search optimizer for solving equation 11. The cal-

culated angles from the previous method from chap-

ter 3.3 can be fed into the optimizer as an initial guess,

which leads to faster convergence. Figure 5 shows the

result of the optimization of a sample data set. The

data set was generated by random acceleration of a

slightly misaligned sensor setup.

The whole alignment procedure is done only once

during startup of the system. Afterwards the rotation

matrix is stored internally and applied to every new

measurement.

3.4 Adaptive Filter

To eliminate noise outside of the respiration rate spec-

trum, we use an IIR Butterworth bandpass ﬁlter on

the acceleration data. The cutoff frequencies for res-

piration rate ﬁltering were set to F

low

= 0.1 Hz and

high

= 0.8 Hz. With only this bandpass the noise

in our signal is still too large. Therefore we imple-

mented an adaptive ﬁlter which works in the follow-

ing way:

• At ﬁrst the data is ﬁltered through a wide 4 stage

IIR Butterworth bandpass with 0.1 Hz to 0.8 Hz.

• Then a spectrum over the last 1 minute of ﬁltered

data is calculated using a FFT. When the intention

is to analyze respiration with higher variance this

time frame can be made shorter.

• The maximum power frequency f

max

, which is

extracted from the FFT, is used to build a new

narrow 4 stage IIR bandpass with bandwidth

. The high border is then calculated as

= f

max

+ f

while the lower border is calcu-

lated with f

= f

max

− f

. For still activities

we chose f

= 0.25 Hz and for sport activities

= 0.50 Hz. The widths of the bandpasses are

chosen in a way that the high variance of the res-

piration frequency while doing sport is taken into

consideration. In situations with less activity like

sitting or sleeping it can be expected that the vari-

ance of the respiration rate is reasonable lower.

The corresponding IIR coefﬁcients are stored in a

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

Figure 5: Left side: Not optimized acceleration data. Here the coordinate systems are misaligned. Right side: Optimized

acceleration data. The coordinate systems are overlapping.

Figure 6: Filtered difference acceleration signal a

diff

database on the chip and the coefﬁcients which ﬁt

best to the calculated frequency band f

to f

are

then applied.

3.5 Peak Detection

After optimization, rotation and ﬁltering we solely

use the y-axis acceleration data a

, which is in the di-

rection normal to the chest and the back surface. This

is the part of the data with the highest information

value to detect the lifting and inclination of the chest

at respiration.

The ﬁnal difference acceleration data a

diff

is then

calculated by subtracting the back sensor data a

back

from the front sensor data a

front

. This reduces the er-

rors introduced by motion of the subject. The result-

ing data is represented in ﬁgure 6.

We then use an embedded peak detection algo-

rithm to detect maxima in the data. The algorithm

stores the last 18 seconds of the data in a buffer and

checks for peaks by looking at every data point A

If the leftmost neighbor A

i−1

and rightmost neighbor

i+1

are smaller than A

, i.e. A

i−1

≤ A

i+1

, A

a peak candidate. The algorithm then calculates the

peak prominence p

of A

. The peak prominence is

deﬁned as vertical height between the peak and the

nearest minimum belonging to a higher peak. If there

is no higher peak it is deﬁned as vertical height be-

tween the peak and the lowest data point.

If the peak candidate A

has an equal or higher peak

prominence p

than 0.5 times the standard deviation σ

of the data in the buffer, i.e. p

≥ 0.5 · σ, it is stored as

a true peak P

After one second the buffer is shifted by 100 points,

the new 100 data points are stored and the search for a

new peak is repeated. This means the possible update

frequency for respiration rate output of our system is

1 Hz.

3.6 Respiration Rate Calculation

From the position of two peaks in the buffer of the

peak detection method introduced in section 3.5 and

sampling rate F

= 100 Hz, we can calculate the time

between two peaks as can be seen in equation 12.

The resulting respiration rate F

is the inverse of T

i.e. F

− P

(12)

3.7 Heart Rate Detection

For parallel heartbeat detection we use the aligned dif-

ferential acceleration signal with a higher sampling

rate of 400 Hz. In the unﬁltered signal of the y-axis

the pulses are already clearly visible (see ﬁgure 7).

Unfortunately this signal is not very suitable for direct

extraction of the heartbeat.

It turns out that for our measurement system the heart-

beat pulses appear as blocks of several swings with

the frequency between 12 and 28 Hz. For ﬁltering out

the undesired noise we then ﬁlter the signal with a 12-

28 Hz IIR Butterworth bandpass ﬁlter. Afterwards we

take the absolute value of the IIR ﬁltered signal (see

ﬁgure 8).

Finally we smooth the signal with a second order FIR

savitzky golay ﬁlter and window of 49 samples, which

Measuring Respiration and Heart Rate using Two Acceleration Sensors on a Fully Embedded Platform

Figure 7: Raw heartbeat signal from differential signal.

Figure 8: IIR ﬁltered absolute heartbeat signal.

keeps the relative maxima, minima and diversiﬁcation

(see ﬁgure 9). Over the resulting signal we use again

our embedded peak detector to obtain the heart rate.

For ﬁltering out outliers, we compare the mean of the

last 5 valid frequencies with the new calculated fre-

quency. If the new one is more than 50% away from

the previous mean it is discarded. At initialization of

the device, when there are no previous frequencies,

we just take the mean of the ﬁrst 10 frequencies as

the value for comparison. Only after the initialization

the heartbeat frequency will be outputted.

Figure 9: Savitzky golay ﬁltered heartbeat signal.

4 RESULTS

To verify our system we performed several test mea-

surements. Two subjects A and B, outﬁtted with

our system, were standing, walking and running on a

treadmill. The reference respiration rate was derived

from a stationary CPX system (Oxycon Pro Care Fu-

sion, San Diego CA, USA). Optoelectronic ﬂow de-

tection occurred by a ﬂow-volume sensing turbine in-

side a transducer holder ﬁxed on the subjects face-

mask. Prior to testing, the ﬂow-volume sensing tur-

bine was calibrated by a 3 liter syringe. Figure 10

shows the experimental setup.

Figure 10: Experimental setup for evaluation of the system

with subject A (left) and B (right).

To smooth the resulting frequencies we applied a 10

second moving average ﬁlter to the measurement and

reference data. The ﬁrst minute of data was skipped

due to calibration and optimization issues.

The error between the measurement and the refer-

ence is stated both as normalized root mean square

error (NRMSE) and mean percentage error (MPE).

The calculation of the NRMSE is done as shown

in equation 13, where the measurement range is

from F

min

= 0.1 Hz to F

max

= 0.8 Hz for respira-

tion rate measurements and from F

min

= 0.8 Hz to

max

= 3.5 Hz for heart rate measurements. The MPE

is calculated according to the equation 14. All data

presented here are recorded from subject A, since the

T-shirt worn by subject B inﬂuenced the measure-

ments and made them unreliable.

NRMSE

∑

i=1

meas

−F

ref

)

max

− F

min

(13)

MPE

∑

i=1

meas

−F

ref

(14)

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

Figure 11: Respiration rate measurement of our system (black) and the reference (blue) in situation S1. The borders of the

adaptive bandpass are marked in red.

Figure 12: Respiration rate measurement of our system (black) and the reference (blue) in situation S2. The borders of the

adaptive bandpass are marked in red.

4.1 Respiration Rate

To quantify the performance of our system in distinct

use cases we reconstructed three different situations.

Situation One (S1): Ofﬁce work.

The subject was sitting on an ofﬁce chair, per-

forming slight motion which involved moving the

upper body and the arms. The measurement time

was 4 minutes. Figure 11 shows the respiration

rate of our system and the reference.

Situation Two (S2): Walking and jogging.

The subject was ﬁrst walking with 5

for 5

minutes, then jogging with 7

for 4 minutes

on a treadmill. The measurements can be seen in

ﬁgure 12.

Situation Three (S3): Running and stopping.

The subject was running with 15

on a tread-

mill followed by 2 minutes of standing still. This

was then repeated, as can be seen in ﬁgure 13.

Table 1 shows the NRMSE and MPE of our system

measurements compared to the reference in the three

situations.

Table 1: NRMSE and MPE of respiration rate measure-

ment.

Situation NRMSE MPE

S1 1.42 % 4.40 %

S2 5.46 % 8.59 %

S3 3.95 % 4.13 %

4.2 Heart Rate

To quantify the performance of or system at measur-

ing the heart rate we used a Polar H7 heart rate belt

strapped to the chest of the subject as the reference.

For this test the situation was slight movement of the

subject, which included body rotation and walking

across a room for 4 minutes. The heart rate measure-

ment of our system and the reference measurements

can be seen in ﬁgure 14. The NRMSE and PME are

stated in table 2.

Table 2: NRMSE and MPE of heart rate measurement.

Situation: Slight movement

NRMSE 1,74 %

MPE 2,65 %

Measuring Respiration and Heart Rate using Two Acceleration Sensors on a Fully Embedded Platform

Figure 13: Respiration rate measurement of our system (black) and the reference (blue) in situation S3. The borders of the

adaptive bandpass are marked in red.

Figure 14: Heart rate measurement of our system (blue) and the reference (red).

5 CONCLUSIONS

In this paper, a method to measure respiration and

heart rate with acceleration sensors has been pre-

sented. The results show that we are able to detect

the respiration rate with low errors in different situa-

tions. By using the differential measurement with two

sensors, advanced calibration, optimization methods,

and an adaptive, rather small bandpass, we are able

to reduce the noise drastically and receive reliable re-

sults, even in very noisy settings such as fast running.

Important to recognize is, that the highest error

occurs if the person has a medium activity like in sit-

uation S2. In this situation the signal-to-noise ratio

is very weak due to the relatively small demand of

oxygen and the corresponding small amplitude of the

chest. Surprisingly the ratio increases again with a

more demanding activity and higher amplitude of the

chest, although this activity results in a higher motion

noise, like in situation S3. This observation explains

the fact that smaller errors can be achieved in very still

activities as well as in intensive sport activities. In

these situations our dual sensor setup produces highly

accurate measurements, while the results in situation

S2 are still very reasonable.

Since it occurred to us that most of the informa-

tion came from the inclination of the sensor during

respiration, further investigations could be done with

gyroscopes in comparison to acceleration sensors.

Furthermore, by using the same sensor setup and sim-

ilar methods we were able to measure the heart rate

with a very small error. Further development of the

heart rate detection component and an application of

our methods to more different situations could be a

promising topic for future investigations.

REFERENCES

Bates, A., Ling, M., Mann, J., and Arvind, D. (2010). Res-

piratory rate and ﬂow waveform estimation from tri-

axial accelerometer data. In Body Sensor Networks

(BSN), 2010 International Conference on, pages 144–

150.

Cao, Z., Zhu, R., and Que, R.-Y. (2012). A wireless portable

system with microsensors for monitoring respiratory

diseases. In Biomedical Engineering, IEEE Transac-

tions on, volume 59, pages 3110–3116.

Carey, D. G., Schwarz, L. A., Pliego, G. J., and Raymond,

R. L. (2005). Respiratory Rate is a Valid and Reliable

Marker for the Anaerobic Threshold: Implications for

Measuring Change in Fitness. volume 4, pages 482–

488.

Heyde, C., Leutheuser, H., Eskoﬁer, B., Roecker, K.,

and Gollhofer, A. (2014). Respiratory inductance

plethysmography-a rationale for validity during exer-

cise. volume 46, pages 488–495.

Hoeﬂinger, F., Mueller, J., Toerk, M., Reindl, L., and Bur-

gard, W. (2012a). A wireless micro inertial measure-

icSPORTS 2015 - International Congress on Sport Sciences Research and Technology Support

ment unit (IMU). In Instrumentation and Measure-

ment Technology Conference (I2MTC), 2012 IEEE In-

ternational, pages 2578–2583.

Hoeﬂinger, F., Toerk, M., Ojha, A., and Reindl, L. (2011).

Drivability enhancement of transport buses using an

auxiliary system with an IMU. International Journal

of Electrical & Computer Sciences (IJECS).

Hoeﬂinger, F., Zhang, R., and Reindl, L. (2012b). Indoor-

localization system using a micro-inertial measure-

ment unit (IMU). In European Frequency and Time

Forum (EFTF), 2012, pages 443–447.

Jin, A., Yin, B., Morren, G., Duric, H., and Aarts, R. (2009).

Performance evaluation of a tri-axial accelerometry-

based respiration monitoring for ambient assisted liv-

ing. In Engineering in Medicine and Biology Society,

2009. EMBC 2009. Annual International Conference

of the IEEE, pages 5677–5680.

Liu, G. Z., Guo, Y. W., Zhu, Q. S., Huang, B. Y., and Wang,

L. (2011). Estimation of respiration rate from three-

dimensional acceleration data based on body sensor

network. In Telemed J E Health, volume 17, pages

705–711.

Onorati, P., Martolini, D., Valli, G., Laveneziana, P.,

Marinelli, P., Angelici, E., and Palange, P. (2012). A

simpliﬁed approach for the estimation of the ventila-

tory compensation point. volume 44, pages 716–724.

Phan, D., Bonnet, S., Guillemaud, R., Castelli, E., and

Pham Thi, N. (2008). Estimation of respiratory wave-

form and heart rate using an accelerometer. In Engi-

neering in Medicine and Biology Society, 2008. EMBS

2008. 30th Annual International Conference of the

IEEE, pages 4916–4919.

Simon, N., Bordoy, J., Hoeﬂinger, F., Wendeberg, J.,

Schink, M., Tannhaeuser, R., Reindl, L., and Schin-

delhauer, C. (2015). Indoor localization system

for emergency responders with ultra low-power ra-

dio landmarks. In International Instrumentation and

Measurement Technology Conference (I2MTC).

STMicroelectronics (2014). AN4508 Application note - Pa-

rameters and calibration of a low-g 3-axis accelerom-

eter.

Tewel, N. (2010). Application of mems accelerometer

for baby apnea monitoring under home conditions.

In Acta Bio-Optica et Informatica Medica. Inynieria

Biomedyczna, volume Vol. 16, nr 4, pages 389–393.

Wasserman, K. (2012). Principles of exercise testing and

interpretation: Including pathophysiology and clini-

cal applications. Wolters Kluwer/Lippincott Williams

& Wilkins, Philadelphia, 5th edition.

Yoon, J., Noh, Y., Kwon, Y., Kim, W., and Yoon, H.

(2014). Improvement of dynamic respiration monitor-

ing through sensor fusion of accelerometer and gyro-

sensor. Journal of Electrical Engineering and Tech-

nology, 1(1).

Zhang, R., Bannoura, A., Hoeﬂinger, F., Reindl, L., and

Schindelhauer, C. (2013). Indoor localization using

a smart phone. In Sensors Applications Symposium

(SAS), 2013 IEEE, pages 38–42.

Measuring Respiration and Heart Rate using Two Acceleration Sensors on a Fully Embedded Platform