Vol. 121 (2012) ACTA PHYSICA POLONICA A No.

Proceedings of the the Tenth Symposium of Magnetic Measurements, Warsaw, 1719.10.2011

Improvement of Electronic Compass Accuracy Based on

Magnetometer and Accelerometer Calibration ∗
M. ’ipo² , J. Rohá£, P. Nová£ek
Czech Technical University in Prague, Faculty of Electrical Engineering, Department of Measurement
Technická 2, 166 27, Prague, Czech Republic

This paper describes the process used for an electronic compass compensation according to accelerometer
based tilt evaluation. Tilt angles have to be estimated rst for sensed magnetic vector components to be aligned
and horizontal components evaluated. Therefore the precision of accelerometer based tilt angles plays a key role in
this whole process as well as the magnetometer characteristics. Hence accelerometers plus magnetometers have to
be calibrated to improve the accuracy of a tilt and an azimuth angle evaluation. The calibration uses Thin-Shell
method to determine sensor error models. Both the eect of calibration and precision of estimated error models
have been observed and are presented. The electronic compass consisted of tri-axial magnetometer and tri-axial
accelerometer contained in the Inertial Measurement Unit ADIS16405 from Analog Devices manufacturer.
PACS: 85.75.Ss, 91.10.v, 06.20.fb, 91.25.r, 07.07.Df

1. Introduction calibration method is mentioned in section 4. A measure-

ment setup and a measured unit are briey introduced in
Since 1500 years ago, the mechanical compasses have
section 5. The most important results are summarized
been used for an azimuth determination and a guidance
in section 6.
using Earth magnetic eld. Due to the technology devel-
opment and improvement, current electronic compasses
(ECs) have much better parameters which are, of course,
2. Electronic compass
inuenced by sensor type applied. The most simple low
accuracy compasses use Hall sensors. In contrast, more
The simplest electronic compass (EC) can be con-
accurate ones use Anisotropic Magneto Resistors (AMR)
structed using only a dual-axis magnetometer. This type
and the most accurate compasses use the uxgate sensors
of EC can measure accurate only azimuth (yaw angle) in
[1]. The nal accuracy of EC depends not only on used
magnetic sensors, but also on tilt sensors, which have
horizontal plane. The resulting azimuth ψ can be com-
puted using simple eq. (1):
to be utilized to mathematically align magnetic sensors ( / )
(compasses with tilt compensation) into the local naviga- ψ = arctan fy fx − D (1)
tion frame. Characteristics of tilt sensors also aect the where fx , fy are horizontal magnetic eld components
EC accuracy, and therefore they have to be calibrated, measured in sensor (body) frame, and D is a magnetic
which eliminates the sensors imperfections [2]. For low- declination [2].
cost sensors like MEMS (Micro-Electro-Mechanical Sys-
Although this type of compasses is very simple and
tem) based ones, manufacturers mostly perform only ba-
easy to manufacture, a main disadvantage of this EC
sic calibration and the rest is left on customers. Thus, for
construction is in the obligation to place the sensor ac-
better accuracy the system needs to be recalibrated [3].
curately into horizontal plane. If it cannot be ensured,
There exists a wide range of calibration procedures and
the errors are not negligible as was proved by Vcelak in
techniques, e.g. the calibration using redundant heading
[2]. Generally, it is not possible to ensure this condition
information computed from rate gyroscopes [4] or the
providing horizontal mounting of magnetic sensor, so the
calibration procedure based on ellipsoid tting problem
electronic compass has to be equipped with tilt compen-
which does not need heading reference information ob-
sation functionality. The compass with tilt compensation
tained from redundant sensors [5].
(Fig. 1) usually consists of tri-axial magnetometer and
Nowadays, the ECs have become useful in a wide range
tilt sensor, which can be formed by tri-axial accelerom-
of consumer applications such as mobile phones, PDAs,
eter [9] or an electronic inclinometer commonly used in
robot navigation, human head and hands tracking, atti-
Honeywell compasses.
tude determination of inertial navigation systems used in
The EC uses magnetometer platform mathematically
aerospace engineering, etc. [1, 2, 68].
aligned to the horizontal plane using pitch and roll an-
In this paper, the EC system and the tilt compensation
gles dened by (2) and (3). The azimuth can be then
is briey described in section 2, the sensor error model
computed using (4).
(SEM) is further discussed in chapter 3 and Thin-Shell ( /√ )
2 2
θ = arctan −a by bx bz
(a ) + (a ) , (2)

( / )
∗ corresponding author; e-mail: φ = arctan abx (−abz ) , (3)
siposmar@fel.cvut.cz

(945)
946 M. ’ipo², J. Rohá£, P. Nová£ek

4. Calibration procedure

There already exist several calibration procedures for

tri-axial sensors using dierent principles, e.g. the
method using an ellipsoidal-tting procedure [5, 13], a
calibration procedure which uses a robotic arm [14] or
a procedure with the usage of 3D optical tracking sys-
tem that measures the position coordinates of markers
attached to a measurement unit [15].
In our case, we used the thin-shell (TS) calibra-
Fig. 1. The block scheme of an EC with a tilt com- tion method. A fundamental principle of the proposed
pensation and the compensation of sensor imperfections method is based on the fact that the magnitude of mea-
formed in the sensor error model (5).
sured quantity |y| (gravity acceleration, magnetic eld
vector) should be always equal to the constant value when
static conditions are ensured and also equal to the square
ψ= (4) root of the sum of squared vector components (6):
fy cos φ + fz sin φ 2
yx2 + yy2 + yz2 = |y| ,
arctan − D, (6)
fx cos φ + fy sin φ sin θ − fz cos φ sin θ where yi denotes sensed quantity in direction of i axis and
bx by
where: θ , φ denote the pitch and roll angle; a , a ,
|y| is the magnitude of measured quantity. In the case of
bz the gravity vector, it is ideally equal to lg and in the case
a are measured accelerations in the sensor body frame;
of the magnetic eld vector |F | it is equal to 0.48125G for
fx , fy , fz represent magnetic eld vector components
the location (area) where the measurements were taken.
measured in sensor frame [10].
The value of Earth magnetic eld vector was calculated
using International Geomagnetic Reference Field model
3. Sensor error model
(IGRF 11) which depends on the date of measurement,
In the chapter 1 it was mentioned that the calibration GPS position, and the altitude [16].
is necessary to be performed for the elimination of the For the calibration purposes, according to [11], 36 po-
sensor imperfections. Generally, sensors have many error sitions are recommended to measure, 3 times 12 posi-
sources; nevertheless, our sensor error model (SEM) de- tions along x, y , z axis. The advantage of the method
ned by (5) includes the main ones [11]. They correspond is that the precise knowledge of position orientations is
to scale factor deections, axes misalignment described not required. It is only recommended to provide at least
in our case by three non-orthogonality angles [11], and 3 positions per each quadrant and each axis. After the
osets for all three axes. The oset forms a stochastic measurements are taken, the Thin-Shell algorithm can
time-invariant part of the bias; in contrast, a drift char- be applied on the measured data. The TS algorithm is
acterizes a time-variant part of the bias. Because the based on a lineaer minimum mean square error princi-
calibration process is commonly performed during short- ple minimizing the standard deviation σ dened by (7),
time period, the drift can be considered as zero [11]. which is calculated from compensated vector component
  estimates and the known number of measurements.
1 0 0 v
  u∑ ( 2 )2
yp = Tap SFa (ym − ba )  αyx 1 0  u m ⌢ ⌢2 ⌢2 2
t i=1 y xi + y yi + y zi − |y|
αzx αzy 1 σ= , (7)
m−1
      ⌢ ⌢ ⌢
SFax 0 0 ymx bax where y xi , y yi , y zi are estimations of compensated accel-
      eration/magnetic eld vector components and |y| is the
×  0 SFay 0   ymy  −  bay  ,(5)
magnitude of the reference value corresponding to mea-
0 0 SFaz ymz baz
sured quantity. The more detailed description of this
where yp represents the compensated vector of either calibration method is presented in [11, 17].
measured acceleration in the case of accelerometers or In each iteration step the interval denes the mini-
magnetic eld vector in case of magnetometers and mum, maximum, and mean value of the parameter being
p
is dened in the orthogonal platform frame; Ta de- searched for and these values are then used to update the
notes the matrix providing the transformation from the SEM. Thus, 3 SEMs are obtained coresponding to min.,
non-orthogonal frame to the orthogonal one with non- max., and mean values of the given parameter. Based on
diagonal terms αyx , αzx , αzy that correspond to the the updated SEMs new estimates of compensated vector
axes misalignment; SFa represents a scale factor ma- are determined for each position and used for σ calcula-
T
trix; ba = [bax , bay , baz ] is the vector of osets; ym = tions. With respect to obtained 3 values of σ the interval
T
[ymx , ymy , ymz ] denotes the vector of measured acceler- is halved to nd the local minimum of a standard devia-
ation/magnetic eld vector. The SEM and its derivation tion according to Fig. 2. When σmean reaches the smallest
are described in more detail in [12]. value, the interval is halved around kmean , where  k  rep-
 Improvement of Electronic Compass Accuracy Based on Magnetometer and Accelerometer Calibration 947

resents the parameter being searched for. Unlikely, when

other σ reaches the smallest value and σmean is the sec-
ond, the new interval is dened between k, whose σ was
the smallest and kmean . For instance, when kmin has the
smallest σ, thekmean becomes kmax for another iteration
step and the new kmean is calculated as the average of
the new kmax and previous kmin . The same principle can
be applied for the other case [11].

Fig. 3. The Inertial measurement unit ADIS16405 (on

the left); theodolite T1c (in the middle); the whole mea-
surement setup (on the right).

parameters of MAG and ACC SEMs are listed in Ta-

ble I. The deviation between the measured and the ideal
vector of applied quantity (corresponds to the magnetic
Fig. 2. Criterions for halving the interval, in which the eld vector for MAG and to the gravity vector for ACC)
estimated parameters are searched for [11]. is shown in Fig. 4 and Fig. 5. In contrast with the chapter
4, in which 36 positions are recommended for a correct
calibration, we used only 21 positions in the case of MAG.
5. Measurement setup The measurement took shorter time and thus we mini-
mized the risk of potential magnetic eld variations. In
In our case the measurement setup was built up by [11] it was proven that 21 positions is a sucient number
the inertial measurement unit (IMU) ADIS16405 [18] without a nal accuracy decrease. For ACC calibration,
(Analog Devices) and the non-magnetic theodolite T1c the 36 positions were measured as was recommended.
(Meopta Prague, Czech Republic). The IMU was used
to evaluate the EC algorithm with tilt compensation and TABLE I
to prove the improvement of applied calibration proce- Sensor error models obtained using Thin-Shell algorithm
dure. The IMU (Fig. 3) contains the tri-axial magne- for magnetometer (MAG) and accelerometer (ACC) of
tometer (MAG), tri-axial accelerometer (ACC), and tri- IMU ADIS16405 (Superscript 1 denotes RMSE before
axial angular rate sensor (ARS). The measurements were
calibration and 2 after calibration)
performed in the area with minimal magnetic eld distur- Parameter MAG ACC
bances in the local time from 18:00 to 19:00 CET when
αxy [deg] 0.1355 0.0230
the variations of magnetic eld are minimal. For the
αzx [deg] 0.6628 0.0351
evaluation of EC accuracy, the IMU was mounted on the
αzy [deg] 0.0818 0.1639
non-magnetic theodolite, see Fig. 3, which was used as a
−3 SFx [] 1.0049 0.9996
reference with an average error 4.17 × 10 deg.
SFy [] 1.0050 1.0019
In all performed experiments we used for calibration
SFz [] 1.0004 0.9983
purposes and a nal EC evaluation the average of 100
ACC and MAG samples taken in each position under
bx 0.71 mG 13.54 mg
static conditions as a value we consequently calculated
by 0.83 mG 6.71 mg
with. A main reason for the usage of average values was
bz 0.23 mG 4.02 mg
the elimination of a noise inuence. RMSE1 1.9 mG 9.5 mg
RMSE2 0.4 mG 2.5 mg
6. Results

6.1. Calibration of Magnetometer and Accelerometer of

IMU ADIS16405 6.2. Inuence of MAG and ACC Calibration to
Electronic Compass Accuracy
From the output data provided by IMU ADIS16405
we used only information from the magnetometer (MAG) Finally, we analyzed in previous chapter performed
and the accelerometer (ACC). After the data had been calibration from the nal accuracy of realized electronic
preprocessed, the calibration was performed using the compass (EC) point of view. We performed four mea-
Thin-Shell algorithm to estimate three misalignment an- surements at all. In each measurement the EC was dif-
gles (non-orthogonality angles), three scale factor cor- ferently tilted in two directions to set values of 0 deg and
rections, and three biases, all formed in SEM (5). The 20 deg in various combinations. Then, the azimuth was
948 M. ’ipo², J. Rohá£, P. Nová£ek

The EC performance generally depends on used tri-axial

magnetometer (MAG) and its parameters as well as on
parameters of an aligning system. In our case we used
tri-axial accelerometer (ACC) for this purpose. To im-
prove EC performance we applied a calibration procedure
Thin-Shell to estimate sensor error models of MAG and
ACC. The methods were shortly introduced; neverthe-
less, a main focus was pointed to present experimental
results. We performed the calibration of MAG, which
approximately ve-times improved its accuracy and in
the case of ACC the accuracy was four-times improved.
Although the calibration procedure recommended 36 po-
Fig. 4. The dependence of deviations of measured
magnetic eld vector before and after calibration  sitions to use, we measured data only 21 in the case of
MAG of ADIS16405  21 evaluated positions. MAG which was in accordance to [11]. In contrast, for
the ACC calibration we kept 36 positions as was recom-
mended. We analyzed the inuence of MAG and ACC
calibration on the nal EC accuracy by analyzing the
dierences between the evaluated azimuth and the ref-
erence angle obtained from our reference system formed
by theodolite T1c. The evaluated azimuth reected esti-
mated SEMs' parameters, which were: three scale-factors
corrections, three non-orthogonality angles, and three o-
sets. In all tested experiments the application of MAG
and ACC SEMs led to improvement of nal EC accuracy
as was presented.

Acknowledgement
Fig. 5. The dependence of deviations of measured ac-
celerations before and after calibration  ACC of This research has been partially supported by Czech
ADIS16405  36 evaluated positions.
Science Foundation project 102/09/H082, partially by
the research program No. MSM6840770015 "Research

changing with the step of 22.5 deg and a tilt compen- of Methods and Systems for Measurement of Physical

sation observed as well as the eect of MAG and ACC Quantities and Measured Data Processing" of the CTU

calibration on the azimuth accuracy. As a criterion for in Prague sponsored by the Ministry of Education, Youth

the azimuth accuracy evaluation the RMSEs were com- and Sports of the Czech Republic and partially by Grant

puted and the nal values with and without calibration Agency of the Czech Technical University in Prague grant

(applied SEM) summarized, see Table II. The table pro- No. SGS10/288/OHK3/3T/13.

vides the nal RMSEs depending on set tilts in two di-

rection (pitch and roll angles). In all four data sets, the References

application of evaluated SEMs led to the improvement of

[1] R. Racz, Ch. Schott, S. Huber, Proceedings of the
the nal EC accuracy.
IEEE sensors 2004, Vienna, Austria 1-3, 1446,
(2005).
TABLE II
[2] J. Vcelak, V. Petrucha, P. Kaspar, Sensor Letters 5,
Final accuracy of yaw angle estimation 279, (2005).
with and without applied ACC and MAG
SEM; θ - pitch, φ - roll, ψ - yaw [3] Z. Syed, P. Aggarwal, C. Goodall, X. Niu, N. El-
Sheimy, Measurement Science & Technology 18, 1897,
θ [deg] φ [deg] without with (2007).
calibration ∆ψRMSE [deg] [4] B. Ho, R. Azuma, IEEE and ACM International
0 0 1.663 0.534 Symposium on Augmented Reality, Munich, Germany,
159, (2000).
0 20 1.270 0.462
20 0 2.012 0.567 [5] X. Hu, Y. Liu, Y. Wang, Y. Hu, D. Yan, 4th
IEEE/ACM International Symposium on Mixed and
20 20 1.303 0.563 Augmented Reality 2005, Vienna, Austria, 182-183,
(2005).
[6] V.Y. Skvortzov, HK. Lee, S.W. Bang, YB. Lee, Pro-
7. Conclusion ceedings of the 2007 IEEE International Conference
on Robotics and Automation 1-10, 2963, (2007).
This paper deals with an electronic compass (EC) algo- [7] M. Reinstein, Przegl¡d Elektrotechniczny 87, 255,
rithm and procedures needed for its correct functionality. (2011).
 Improvement of Electronic Compass Accuracy Based on Magnetometer and Accelerometer Calibration 949

[8] M. Reinstein, M. Sipos, J. Rohac, Przegl¡d Elek- [16] NOAA's Geophysical Data Center - Ge-
trotechniczny 85, 114, (2009). omagnetic Online Calculator. Available:
[9] J. V£elák, P. Ripka, J. Kubík, A. Platil, P. Ka²par, http://www.ngdc.noaa.gov/ geomagmodels/ IGR-
Sensors and Actuators A: Physical 123-124,, 122, FWMM.jsp [Accessed: 09-Sep-2011].
(2005). [17] M. Soták, M. Sopata, R. Bréda, J. Rohá£, L. Váci,
[10] M. Soták, Przegl¡d Elektrotechniczny 86, 247, (2010). Navigation System Integration, Ko²ice: Robert Breda,
Kosice, the Slovak Republic (2006).
[11] M. Sipos, P. Paces, J. Rohac, P. Novacek, IEEE Sen-
sors Journal, (2011), accepted for publication. [18] ADIS16405 High Precision Tri-Axis Gyro-
[12] I. Skog, P. Händel, XVII IMEKO World Congress,
scope, Accelerometer, Magnetometer, Inertial
Rio de Janeiro, (2006).
Sensors, Sensors, Analog Devices. Available:
http://www.analog.com/en/sensors/inertial-
[13] S. Bonnet, C. Bassompierre, C. Godin, S. Lesecq, sensors/adis16405/products/product.html [Accessed:
A. Barraud, Sensors and Actuators A: Physical 156, 24-Apr-2011].
302, (2009).
[14] E.L. Renk, M. Rizzo, W. Collins, F. Lee, D.S. Bern-
stein, IEEE Control Systems Magazine, 25, 86,
(2005).
[15] A. Kim, MF. Golnaraghi, Plans 2004: Position loca-
tion and Navigation Symposium, Monterey, CA, 26,
(2004).