5 th Mediterranean Conference on Embedded Computing ",

111
1 , MEeO 2016 Bar, Montenegro

Exploiting Accelerometers to Estimate Displacement

Renato Ferrero, Filippo Gandino, Masoud Hemmatpour, Bartolomeo Montrucchio, Maurizio Rebaudengo
Dipartimento di Automatica e Informatica
Politecnico di Torino
Torino, Italy
{renato.ferrero, filippo.gandino, masoud.hemmatpour, bartolomeo.montrucchio, maurizio.rebaudengo}@polito.it

Abstract- Although the acceleration is physically related to the systems for studying and monitoring human actlvity and
displacement of an object, i.e., to its change of position, it is mobility [5]. Advances in integrated MEMS further reduced
demonstrated that the double integration of the acceleration does size and cost of accelerometers, while ensuring their high
not provide accurate information about the displacement, due to quality and reliability as required by industrial standards [9].
the noise and measurement errors. This paper evaluates a
correction technique based on the Kalman filter in order to This paper analyzes the motion through a single
increase the accuracy of the estimation of the displacement. accelerometer. The goal is to track the movement, i.e., to detect
Experiments were performed by acquiring the acceleration with when an object changes its position and to compute its
an off-the-shelf accelerometer: the percentage error made by displacement from the initial location, by means of the data
simply integrating the acceleration measurements may arrive to collected by the accelerometer. The basic approach relies on
68% in the general case of a movement in the space, but it can be the double integration of the accelerometer data: the first
dramatically reduced to 9% with the proposed approach. An integration retrieves the current speed and then the second
even better behavior is obtained when the movement is integration computed on the speed provides the displacement.
constrained to a plane or along an axis. The innate problem in that approach is due to the noise that
affects the measurements. In this paper, this issue is managed
Keywords- accelerometer, Kalman jilter, position tracking, by implementing the Kalman filter [10, I I], commonly adopted
displacement
in many application fields, like in navigation and control
I. INTRODUCTION systems. The experiment was performed by using a Waspmote,
a wireless sensor platform commercialized by Libelium with an
Accelerometers are sensors for measuring the acceleration embedded three-axes linear accelerometer [12]. The three-axes
of objects in motion along reference axes [1]. An accelerometer accelerometer measures the acceleration of an object in motion
consists of a mass connected to the case by means of a beam, along three reference axes: in this way, different case studies
which can be represented with a damped spring. The are considered, by studying the motion along one axis, in a
acceleration is measured by the deformation of the spring, plane and in the space. The results of the experiment reveal
which is converted into a change of the impedance (in a that, by means of the Kalman filter, the developed technique
resistive or capacitive accelerometer) or into an electrical notably increases the accuracy of the estimation of the
signal (in a piezoelectric accelerometer). displacement. Without any correction technique, the average
Measuring acceleration is a proper approach for analyzing percentage error between the estimated and the real value, as
motion and displacement (i.e., change of position) because obtained with a measuring tape, ranges between 44.8% and
acceleration is proportional to the external force and hence it 68.0%, depending on the considered case study. However, this
can reflect the intensity and frequency of movement [2]. error dramatically drops, ranging between 3.4% and 9.2%, with
Information about speed and displacement can be obtained by the adopted implementation ofthe Kalman filter.
integrating accelerometer data with respect to time. In addition, The remainder of the paper is organized as follows. Section
it is possible to retrieve some information about inclination, II reviews state-of-the-art solutions to estimate a displacement
like the tilt of an object rotating with respect to a reference based on data acquired from an accelerometer. Section III
plane and its orientation [3]. Besides, accelerometers have been formulates theoretically the method adopted in this study and
widely exploited in wearable devices to measure and assess Section IV discusses its accuracy by providing experimental
various kinds of human activities in either clinical/laboratory results. Finally, some conclusions are drawn in Section V.
context or free-living environments [4, 5]. For instance, the
first use of accelerometers to measure human motion dates II. RELATED WORK
back to the 1950s [6]. Technological advances in the 1970s
Retrieving information about the position by integrating the
expanded the use of accelerometers [7]: the micro-electro-
signal acquired from an accelerometer is a well-established
mechanical system (MEMS) technology greatly reduced both
method [13]. However, the naive algorithm can be applied only
device cost and power consumption, and enhanced sensor
if precision is not critical because the acquired signal contains
performance. The first batch-fabricated MEMS accelerometers
noise that should be properly processed. A correction technique
date to 1979 [8], then many research and commercial
needs to be implemented, according to the application scenario.
applications integrated MEMS accelerometers into wearable

This work has been partially sponsored by Project OPLON (Opportunities for active and healthy longevity), Call "Smart Cities and Communities", funded by
Ministero Istruzione, Universita e Ricerca, Italy.

-206-
 5 th Mediterranean Conference on Embedded Computing ",
111
1 , MEeO 2016 Bar, Montenegro

N-l
The tracking of human movement frequently exploits one
"\"' _; ZIT"r
or more inertial and magnetic sensors attached to the A" = L a"e N (I)
significant parts of the body, like limbs and feet [14]. Many "=0
types of human movements, like walking, running, and side where i is the imaginary unit. The discrete Fourier transform
stepping, are composed of cyclic phases. For instance, the of the displacement directly derives from the properties of the
stance and the swing phases alternate during walking. There Fourier transform of the integrals:
can be some periodic moments with instant speed equal to 1
zero: this occurs for example at the transition between stance D" = - C2rrk)Z A" . (2)
and swing phase, when both the feet are in contact with the
Finally, the discrete series of the displacement is obtained by
ground. The recognition of these moments can improve the
applying the inverse Fourier transform to the coefficients in (2)
accuracy of the measurements. If the expected speed is zero but
the value obtained by integrating the acceleration is different, N-l
"\"' . ZIT
the error is due to drift in the measurements and it can be used dr = L D"elN"r . (3)
to correct the accelerometer data, in order to obtain a consistent
"=0
speed value. In this way, a more accurate information about the The scale factor of the error involved in (3) is -l/(2nk)2. This
human movement is obtained by integrating the speed signal: means that the displacement level drops off at a rate of 6 dB as
e.g., the average error in indoor walking experiments is 5.5%. the frequency doubles, so the dominant signal components can
be lower than the low-frequency noise components during the
Inertial-based measurement techniques are common for
conversion process. In order to prevent the low-frequency
estimating the displacement in automotive engineering [15].
components from being significant, it is suggested to zero
First, accelerometers are small devices and they can be easily
padding the frequency components except the significant
attached to the vehicle without loading it down. Secondly, the
sinusoidal components appearing in the Fourier transform of
accelerometer data are useful to study the structural integrity of
the acceleration signal [17]. A more precise method requires
the vehicle: the damage to the structure depends on the stress,
the extraction of the significant frequency components in the
which is proportional to the displacement, for an elastic
Fourier transform of the acceleration signal before it is inverse-
structure. Finally, the wide frequency and dynamic range of the
Fourier transformed [18]. This task can be accurately
accelerometers facilitate the study of vibrations. The vibration
performed by means of curve fitting around the peak values of
pattern and the movement of the vehicle on the road are mainly
the discrete Fourier transform. Curve fitting is the process of
characterized by low-frequency random acceleration data. With
finding the coefficients in a mathematical function that most
these predominant bands of frequencies, three approaches of
closely matches a series of data points.
digital filtering based on finite impulse response, infinite
impulse response, and the inverse of the fast Fourier transform Besides the error introduced by integrating the acceleration
are investigated to reduce integration errors [15]. The standard signal, the estimation of the displacement is affected also by
error increases as the frequency of the acceleration data grows: the double integration of thermal noise, drift and any other
it ranges from a minimum of 5.7% at 20 Hz to a maximum of noise effect [19, 20]. The latter contribution is easily revealed
21.5% at 150 Hz. The accuracy ofthe three filters is similar. by taking measurements from a stationary accelerometer: its
root mean square value increases with the integration time and
The estimation of the displacement based on the
depends also on the sampling frequency. A theoretical model
acceleration measurements is actively investigated in structural
of this error, considering the effects of both analytical [19] and
health monitoring for damage detection [16]. In this scenario,
numerical [20] integration, can be of help for a preliminary
the on-line computation of the displacement can be sacrificed
calibration of the accelerometer.
for a better estimation. Thus, the noise amplification due to the
integration of the acceleration signal can be limited by means III. THEORETICAL APPROACH
of an improved Kahnan filter. In particular, in off-line data
processing the entire sequence of measurements can be used to Given an object that is stationary at the initial time, its
estimate the condition at a specific point, so allowing the speed at time T is given by the integral of the acceleration over
Kalman filter smoothing. The smoothing works through a time. By the definition of integral, the speed can be expressed
combination of forward and backward Kalman filtering over as a summation:
the whole sequence of samples. Moreover, as different
sampling rates are commonly used in a civil engineering
application, a multi-rate Kalman filter should be applied. The
error of the displacement depends on the acceleration data: e.g.,
vCT) = i
T

aCt)dt =
n
~~n;, ~ ~ aCts) . (4)

As expressed in the right part of (4), the speed can be

with a swept-sine signal, the root mean square error is 2.5%. computed as the sum of all small areas below the acceleration
Converting the acceleration signal in the frequency domain curve in the Cartesian plane, according to the trapezoidal rule.
is an alternative approach to the integration in the time domain This is the approach followed in Algorithm 1, which is
in order to compute the displacement [17, 18]. The discrete implemented in the experiments to estimate the speed.
series {ak} composed of N acceleration samples is converted in Algorithm 1 expects as input the value of the acceleration
the frequency domain through the discrete Fourier transform: along an axis and it gives as output the value ofthe speed along

-207 -
 5 th Mediterranean Conference on Embedded Computing ",
111
1 , MEeO 2016 Bar, Montenegro

input, B is an n x I matrix that represents the control-input

1: speed_initialx = speed_new, II the initial speed is model applied to Uk, and Wk is the process noise. At the
II equal to the last computed value instant k, a measurement Zk E IRl.m of the true state Xk is made
2: time = time + ~t according to:
3: accx = ACC.getxO II read the acceleration
4: if (accx < threshold_low) II adjustment Zk=Hxk+Vk (6)
5: accx = threshold_low where H is an m x n matrix that relates the true state Xk to the
6: if (accx > threshold_high) observed state Zk and Vk is the measurement noise. Both the
7: accx = threshold_high process and measurement noises are considered white, i.e., not
8: speed_ newx = speed_initialx + accx * ~t correlated in time, and Gaussian, i.e. their probability density is
equal to that of the normal distribution. Let Q and R be the
Algorithm 1. Computation of the speed along the x-axis matrices of covariance of the process and measurement noise,
respectively.
\: X initial = x new II the initial position is equal Let Xk I k-l E IRl.n be the a priori state estimate at time k,
II to the last computed value i.e., the state estimate that is predicted by knowing only the
2: call Algorithm 1 II computation of the speed process before time k. Similarly, let Xk IkE IRl.n be the a
3: dx = (speed_initialx + speed_ newx) /2 * ~t posteriori state estimate at time k, which exploits also the
4: x new = x initial + dx knowledge of the measurement zk. The a priori and a
posteriori estimate errors are:
Algorithm 2. Computation of the displacement along the x-axis
the same axis. In order to track an object moving in the space, ek I k-l = Xk - Xk I k-l
(8)
Algorithm 1 should be repeated 3 times. Checking the values ek I k = Xk - Xk Ik
acquired by the accelerometer prevents spurious data from and the a priori and a posteriori estimate error covanance
being used. The variables speed_initial, speed_new, and time matrices are:
are initialized to zero. The thresholds differ for each axis and
are related to the measured acceleration: e.g., thresholds for z- P k I k-l = E[ek I k-l ell k-l] (9)
axis are higher than for other axes, because the acceleration P klk = E[eklk el, k ]
along z-axis is the highest one, as it includes gravity force. The Kalman filter uses a feedback control to estimate the
Similarly, the displacement is defined as the integral of the process: firstly, it estimates the process at time k without
speed, but it can also be expressed as a summation, according knowing Zb then it can refine the estimate with Zk. The two
to the approach followed by Algorithm 2: phases are called predict and update, respectively. The predict

i b
phase computes the a priori estimates of the state and error
T n
covariance matrix as follows:
d(T) = v(t)dt = ~~';; ~ vCts) . (5)
Xk I k-l =A Xk-l I k-l +
(10)
BUk
The value of the displacement computed in (5) may Pk I k-l = A Pk - 11 k_1 AT + Q .
significantly differ from the real displacement due to noise and The update phase implements the feedback control by adding
measurements errors. The computational error can be reduced the value of the new measurement to the a priori estimate in
with the Kalman filter, which is an optimal recursive data order to obtain a better a posteriori estimate:
processing algorithm [\ 0]. It is optimal because it estimates the
current value of the parameters by statistically minimizing the Xk I k = Xk I k-l + Kk(Zk - HXk I k-l)
(1\)
error. It is recursive because when a new measurement is P k I k = (I - KkH)Pk I k-l
acquired, previous data are not reprocessed again. It is a data where K, which is called Kalman gain, is a n x m matrix that
processing algorithm because it receives discrete-time minimizes the a posteriori error covariance:
measurement sample in input, it processes them and it provides
an estimation of the parameters of interest in output. Pk I k-l HT
Kk = (12)
The information processed by the Kalman filter includes H Pk I k-l HT + R
the description of the dynamics model of the system, the The adopted implementation of the Kalman filter to
available measurements, and their uncertainty (i.e., process compute the speed along the x-axis is shown in Algorithm 3. In
noise and measurement errors) [11]. In order to be manipulated the predict phase, the a priori estimate of the speed is obtained
more easily, the system model needs to be linear, i.e., it can be by multiplying the a posteriori estimate computed in the
described by a linear stochastic difference equation: previous iteration of Algorithm 3 by a constant A: there are no
control inputs. Similarly, the a priori estimate of the error
Xk = AXk-l + BUk + Wk (5) covariance is based on the a posteriori estimate of the error
where Xk E IRl.n is the state of the system at time k, A is an n x covariance computed in the previous iteration of the algorithm.
n matrix that expresses how the system evolves to the next step The a priori estimate of the error covariance is used in the
without any process noise, Uk E IRl. 1 is the optional control update phase to compute the Kalman gain. The measurement

-208 -
 5 th Mediterranean Conference on Embedded Computing ",
111
1 , MEeO 2016 Bar, Montenegro

TABLE!. DISPLACEMENT ALONG THE X-AXIS

1: priori_speedx = A * posteriori_speedx //predict phase
2: priori_Px = A * posteriori_Px * A_t + Q d, (cm)
acceleration integral with Kalman filter
v(O) = 0
3: z speedx = speed_ newx // update phase dx (em) % error dx (em) % error
4: K= priori_Px * H_t / (H * priori_Px * H_t + R) yes 10 17.9 79.5% 11.7 16.7%
no 75 99.6 32.8% 76.8 2.0%
5: posteriori_speedx = priori_speed, +
yes 10 16.2 61.6% 10.5 5.1%
+ K * (z_speedx - H * priori_speedx ) no 40 47.6 18.9% 41.9 4.8%
6: posteriori_Px = (I - K * H) * priori_Px yes 15 21.5 43.6% 14.0 6.6%
yes 30 48.5 61.6% 32.1 6.9%
yes 10 17.1 70.5% 11.1 10.9%
Algorithm 3. Use of the Kalman filter to compute the speed along the x-axis no 30 50.3 67.5% 33.2 10.7%
Total 220 318.6 44.8% 231.0 5.0%

TABLE II. DISPLACEMENT IN THE XY-PLANE

acceleration integral with Kalman filter

dxy
(cm) dx dy dxy % dx dy dxy %
(em) (em) (em) error (em) (em) (em) error
35 49.3 17.9 52.5 50.1% 32.1 11.7 34.2 2.4%
50 32.3 72.7 79.6 59.1% 21.0 47.3 51.7 3.5%
90 125.6 61.0 139.6 55.1% 83.0 40.3 92.2 2.5%
95 118.4 103.1 157.1 65.3% 78.2 68.1 103.8 9.2%
35 48.5 26.9 55.4 58.4% 31.8 17.8 36.5 4.2%
80 78.1 92.4 121.0 51.2% 51.6 60.8 79.7 0.3%
385 452.2 374.2 605.2 57.2% 297.8 246.1 398.2 3.4%

In the first experiment, the Waspmote moved along a

straight line on a desk. The direction of the line coincides with
the x-axis of the accelerometer embedded in the Waspmote.
The experiment was repeated several times, as shown in Table
Figure I. The Waspmote with the embedded accelerometer used in the tests. I, by varying the distance covered and the initial speed. Amon~
of the speed used in this phase corresponds to the last value the records in Table I, if the speed is different from zero It
computed by Algorithm 1. Then, the a posteriori estima!es of means that the movement is a continuation of that reported in
the speed and error covariance are updated to be used In the the previous row. For example, in the first realization, the
next iteration of Algorithm 3. The pseudocode of the Waspmote started moving (the initial speed is. zero) and
computation of the displacement with the Kalman filter is covered 10 cm. Two different estimations of the displacement
similar to the one in Algorithm 3 and it is not reported here for based on the accelerometer data were provided at this
brevity. checkpoint. Without any break, the Waspmote continued
moving and covered 75 cm more. Two estimations of the
IV. EXPERIMENTAL ANALYSIS partial displacement were computed at the. end o.f the
movement. The estimation based on the double IntegratlOn of
The accelerometer used in the experiment is an
the acceleration is significantly higher than the real
L1S344ALH sensor [21], built-in on the board of a Waspmote,
displacement in all the cases presented in Table I. The
which is shown in Fig. 1. This three-axis linear accelerometer
percentage error ranges between 18.9% and 79.5%; t.he err~r
owns a sensing element and an IC interface. The sensing
made over the distance covered in the whole expenment IS
element includes inertial sensors and actuators in silicon. The 44.8%. The application of the Kalman filter reduces notably
IC interface acquires the acceleration data from the sensing
the value of the estimation, which generally remains higher
element and provides an analog signal in output. The than the real distance but in one case it is slightly lower. The
L1S344ALH accelerometer has a user selectable full-scale of highest percentage error is 16.7% and the average one is 5.0%.
±2 g / ±6 g and it can measure accelerations over a maximum
bandwidth of 1.8 kHz. The programming language for the In the second experiment, the Waspmote freely moved on
Waspmote is C for Arduino. the desk. Table II reports the data collected in distinct
realizations ofthe experiment. After estimating the components
In the experiments, the distance covered by the Waspmote,
d, and dy along the x- and y-axis, the displacement d,y is
obtained with a measuring tape, is used as reference to check computed with the Euclidean norm. The initial speed of the
the accuracy of the estimation based on the data acquired by
Waspmote is not indicated in Table II because it is al:vays.zero.
the accelerometer. Two estimated displacements are compared:
The last row of Table II summarizes the results obtaIned In the
the first one is obtained with the naive double integration ofthe
experiment. Similarly to the motion along the x-axis, the
acceleration and the second one is refined by applying the
distance estimated with the double integration is notably higher
Kalman filter. Three experiments are performed, by moving the
than the real value: the percentage error is always higher than
Waspmote along the x-axis, in the xy-plane and in the space.
50%, and the error made over the whole distance covered in the

-209-
 5 th Mediterranean Conference on Embedded Computing ",
111
1 , MEeO 2016 Bar, Montenegro

experiment is 57.2%. The percentage error dramatically REFERENCES

reduces by applying the Kalman filter: the highest error is 9.2% [I] J. Shieh, J. E. Huber, N. A. Fleck, and M. F. Ashby, 'The selection of
and the average one is 3.4%. The Kalman filter decreases the sensors", Progress in materials science, vol. 46, no. 3, pp. 461-504,
value of the estimated distance, which becomes closer to the 200l.
real value and in one case it is even slightly lower. [2] K. Y. Chen, and D. Bassett, 'The technology of accelerometry-based
activity monitors: current and future", Medicine and science in sports
In the third experiment, the Waspmote moved in the space and exercise, vol. 37, no. 11, pp. 490-500, 2005.
following different paths, as shown in Table III. In the first [3] M. Pedley, "Tilt sensing using a three-axis accelerometer", Freescale
realization, the Waspmote moved without constraints. In the Semiconductor Application Note, pp. 1-22,2013.
next two realizations, the position along the x-axis was kept [4] M. J. Mathie, A. C. F. Coster, N. H. Lovell, and B. G. Celler,
fixed. Similarly, two other movements were performed without "Accelerometry: providing an integrated, practical method for long-
term, ambulatory monitoring of human movement", PhYSiological
varying the position along the y-axis. Finally, the position in
measurement, vol. 25, no. 2, pp. RI-R20, 2004.
the xy-plane was kept fixed and the Waspmote moved only
[5] C.-C. Yang, and Y.-L. Hsu, "A review of accelerometry-based wearable
along the z-axis. In all the realizations of the experiment, the motion detectors for physical activity monitoring", Sensors, vol. 10, no.
distance covered along the z-axis was 60 cm and it was used as 8,pp.7772-7788,2010.
reference to evaluate the accuracy of the estimation of the [6] J. B. Saunders, V. T. Inman, H. D. Eberhart, 'The major determinants in
displacement dz. The data collected in the experiment confirm normal and pathological gait", The journal of bone & joint surgery, vol.
that simply integrating the acceleration does not provide an 35,no.3,pp. 543-558, 1953.
accurate estimation of the distance: the average estimation is [7] J. R. W. Morris, "Accelerometry - a technique for the measurement of
about I m, with an average percentage error of 68.0%. The human body movements", Journal of biomechanics, vol. 6, no. 6, pp
729-736, 1973.
Kalman filter is highly beneficial: the average estimation drops
to 65 cm, with an average percentage error of9.2%. [8] 1. M. Roylance, and J. B. Angell, "A batch-fabricated silicon
accelerometer", IEEE Transactions on Electron Devices, vol. 26, no. 12,
pp. 1911-1917, 1979.
TABLE III. DISPLACEMENT IN THE SPACE
[9] K. M. Culhane, M. O'Connor, D. Lyons, and G. M. Lyons,
acceleration integral with Kalman filter "Accelerometers in rehabilitation medicine for older adults", Age and
dz ageing, vol. 34, no. 6, pp. 556-560,2005.
(cm) dx dv dz % dx dy dz %
(em) (em) (em) errorz (em) (em) (em) errOTz [10] P. S. Maybeck, "Stochastic Models, Estimation, and Control. Vol. 1",
60 45.7 31.4 100.1 66.9% 29.8 20.4 65.1 8.3% Academic press, 1979.
60 0 147.2 103.3 72.2% 0 95.7 67.2 12.0% [II] G. Welch, and G. Bishop, "An introduction to the Kalman filter",
60 0 112.2 103.0 71.7% 0 73.0 67.0 11.7% University of North Carolina at Chapel Hill, pp. 1-16,2006.
60 102.3 0 99.0 64.9% 66.6 0 64.3 7.2% [12] Libelium, www.libelium.com/products/waspmote. accessed 04/0412016.
60 143.6 0 99.2 65.3 93.4 0 64.5 7.5% [13] K. Seifert, and O. Camacho, "Implementing positioning algorithms
100.1 65.1 using accelerometers", Freescale Semiconductor, 2007.
60 0 0 66.8% 0 0 8.5%
360 291. 7 290.8 604.7 68.0% 189.7 189.2 393.3 9.2% [14] X. Yun, and E. R. Bachmann, E. R., H. Moore IV, and J. Calusdian,
"Self:contained position tracking of human movement using small
inertial/magnetic sensor modules", in IEEE International Conference on
V. CONCLUSIONS
Robotics and Automation, pp. 2526-2533, 2007.
The estimation of the displacement based only on the data [15] 1. D. Slifka, "An accelerometer based approach to measuring
acquired from the accelerometer is impractical in most of the displacement of a vehicle body", Master of Science in Engineering,
cases, because the error made in the estimation is too high. As Department of Electrical and Computer Engineering, University of
Michigan-Dearborn, 2004.
a general rule, the estimation is always higher than the real
[16] A. Smyth, and M. Wu, "Multi-rate Kalman filtering for the data fusion
distance and the percentage error between the two values tends of displacement and acceleration response measurements in dynamic
to rise as the freedom of movement of the object increases. The system monitoring", Mechanical Systems and Signal Processing, vol.
average results obtained from experimental data indicate that 21, no. 2, pp. 706-723,2007.
the percentage error ranges between 45%, in the case of a [17] S. Han, S. "Retrieving the time history of displacement from measured
movement along an axis, and 68%, in the case of a movement acceleration signal", KSME international journal, vol. 17, no. 2, pp.
in the space. A correction technique based on the Kalman filter 197 -206, 2003.
can provide significantly better estimation of the displacement: [18] S. Han, "Measuring displacement signal with an accelerometer", Journal
of Mechanical Science and Technology, vol. 24, no. 6, pp. 1329-1335,
the average percentage error drops to 5% if the movement is 2010.
constrained along an axis, and to 9% in the case of a movement
[19] Y. K. Thong, M. S. Woolfson, 1. A. Crowe, B. R. Hayes-Gill, and R. E.
in the space. Therefore, the implemented correction technique Challis, "Dependence of inertial measurements of distance on
can be profitably applied in application scenarios where the accelerometer noise", Measurement Science and Technology, vol. 13,
precision of the displacement estimation is not critical. no. 8, pp. 1163-1172, 2002.
[20] Y. K. Thong, M. S. Woolfson, J. A. Crowe, B. R. Hayes-Gill, and D. A.
ACKNOWLEDGMENT Jones, "Numerical double integration of acceleration measurements in
noise", Measurement, vol. 36, no. 1, pp. 73-92, 2004.
The authors thank the student Melaku Girma Habtewold for
[21] ST Microelectronics, www.st.com/web/enlresource/technical/document/
his support in setting up the experiments during his M. Sc. datasheet/CDOO 182781.pdf, accessed 04/0412016.
thesis.

-210-