You are on page 1of 12

1408 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO.

2, FEBRUARY 2018

Compensated Heading Angles for Outdoor


Mobile Robots in Magnetically
Disturbed Environment
Jehong Lee , Graduate Student Member, IEEE, Jeonggeun Lim, Graduate Student Member, IEEE,
and Jongho Lee , Member, IEEE

Abstract—Heading information is critically important for extensively [1], [2] for daily applications including aerial pho-
autonomous mobile robots as it is necessary for scanning tography and household cleaning [3]–[5]. In addition, new types
or sweeping predetermined areas for specific tasks. Fus- of mobile robots are being actively developed as consumer prod-
ing sensor data including angular rates, acceleration, and
geomagnetic fields provide heading and attitude. However, ucts [6], [7]. These small-scale and low-cost mobile robots have
the geomagnetic field is often interfered with by ferromag- been realized due to advancements of low-cost microelectrome-
netic objects or other magnetic sources, resulting in incor- chanical systems (MEMS) sensors including 3-axis accelerome-
rect heading information. This paper describes an algorithm ters, gyroscopes, and magnetometers [8]. Fusing measurements
that detects and rejects magnetic disturbances contained
from these MEMS sensors provides important data such as at-
in a geomagnetic field. This algorithm combined with an
extended Kalman filter is implemented in a relatively low- titude (roll, pitch, yaw) or heading angles (yaw) that are critical
cost, small-scale microprocessor and sensor module. The for navigation and control of mobile robots. For example, ac-
algorithm is detailed for parameters that detect magnetic curate heading angle estimation of mobile robots is critical for
disturbances. The algorithm is also evaluated outdoors by performance in fine sweeping of predefined areas for tasks such
driving a mobile robot on a lawn with apparent ferromag- as cleaning, vacuuming, and many other potential applications
netic objects and on the flat roof of a ferroconcrete building
that includes iron bars and electrical wires in or under the [9]–[13].
roof. The experimental results on a flat roof indicate that Fusing sensors to estimate attitude or heading angle is usually
the algorithm improves the accuracy of the heading signif- realized by extended Kalman filter (EKF) [14]. Integrating mea-
icantly by reducing the peak-to-peak error by 32.9% (or the sured angle rates from a 3-axis gyroscope predicts roll, pitch, and
rms error by 69.9%). yaw which usually drift because of the accumulation of errors or
Index Terms—Accelerometers, attitude and heading ref- noise from the gyroscope sensors. Correcting attitude regularly
erence system (AHRS), gyroscopes, heading angle, iner-
tial measurement unit, Kalman filter, magnetic disturbance,
via measurement data from other sensors helps to avoid such
magnetometers, mobile robots, navigation, sensor fusion. drifting. A 3-axis accelerometer that provides information for
inclination by reading gravitational acceleration can correct roll
NOMENCLATURE and pitch. Heading angles (yaw) can be corrected by measure-
ẋ Deviation of vector. ments from a 3-axis magnetometer which reads geomagnetic
xk Discretized kth vector. flux density [15]–[19].
x− A priori vector predicted by system dynamic model. Although magnetometers are convenient for correcting head-
x+ A posteriori vector corrected by measurement model of ing angles, the sensors read geomagnetic flux density including
Kalman filter. magnetic disturbances caused by ferromagnetic objects or other
H[1] Jacobian matrix of the matrix H. magnetic sources. A disturbed magnetic field produces errors in
the heading angles. Though calibration methods are suggested
I. INTRODUCTION
to overcome the influence of magnetic objects around magne-
ECENTLY, small-scale mobile robots such as quadro-
R tors or unmanned ground vehicles have been used more
tometers, those methods are only useful at fixed locations [20],
[21]. Other methods that use one or two parameters to detect
magnetic disturbances when fusing different types of sensors
Manuscript received October 11, 2016; revised February 20, 2017 and can be effective, but such results are often from measurements
May 17, 2017; accepted June 30, 2017. Date of publication July 17, 2017; on fixed platforms in confined indoor environment [22]–[25].
date of current version December 8, 2017. This work was supported
in part by the GIST Technology Institute and in part by the National However, outdoor mobile robots may go through magnetically
Research Foundation of Korea under Grant NRF-2016R1A2B4012854. disturbed environment including ferromagnetic pipes and wires
(Corresponding author: Jongho Lee.) or electric wires with current flowing underground.
The authors are with the School of Mechanical Engineering, Gwangju
Institute of Science and Technology, Gwangju 61005, South Korea In this paper, we focus on removing errors in heading angles
(e-mail: kiddhong@gist.ac.kr; vesta8622@gist.ac.kr; jong@gist.ac.kr). for mobile robots in magnetically disturbed environment. We
Color versions of one or more of the figures in this paper are available present an algorithm that detects and rejects magnetic distur-
online at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIE.2017.2726958 bances with details in various situations by combining three
0278-0046 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
LEE et al.: COMPENSATED HEADING ANGLES FOR OUTDOOR MOBILE ROBOTS IN MAGNETICALLY DISTURBED ENVIRONMENT 1409

parameters to detect magnetic disturbances. This algorithm


combined with EKF is implemented in a small-scale, low-cost
microprocessor and sensor module for experimental demon-
stration of the concept. We present detailed experimental results
for the parameters that can detect magnetic disturbances in mag-
netic flux density measured with a 3-axis magnetometer. Finally,
we describe experimental results acquired while driving mobile
robots on a lawn with and without apparent ferromagnetic ob-
jects, on the flat roof of a ferroconcrete building, on a parking
lot, and on a sidewalk. The results presented here improve ac-
curacy of the attitude and heading reference system (AHRS)
in environment where magnetic disturbances exist. In addition,
the experimental results demonstrate the feasibility of the con-
cept by comparing the estimation results with and without the
algorithm implemented in the same EKF. Fig. 1. Configuration of the extended Kalman filter to estimate atti-
tude. The attitude (Ψ − = [Φ − , θ − , Ψ − ]) is predicted through a system
dynamic model with the measured angular rate ω from a gyroscope.
II. KALMAN FILTER IMPLEMENTATION The measured acceleration (A) points down because gravity is used
to correct roll (Φ + ) and pitch (θ + ) through an acceleration measure-
A Kalman filter is an optimal estimator of unknown states ment model. Finally, yaw (Ψ + ), i.e., heading angle, is corrected with the
which fuses measurement data from various sensors that usu- measured magnetic flux density (B) which points magnetic north.
ally contain inaccuracies due to noise or drift [26]. The filter
estimates states by following prediction and correction steps. In intuitive, we use a quaternion as states in a system dynamic
a prediction step, the filter estimates the states based on a sys- model to represent attitude to avoid obstacles such as gimbal
tem dynamic model. In a correction step, the filter corrects the lock [28]:
states based on a measurement model. If the models are non-
linear, the EKF works by using linearized models around the x = q = [q0 , q1 , q2 , q3 ]T . (1)
current estimated state [27]. The AHRS presented here is based The time derivative of the quaternion [29] is related to the
on an EKF whose configuration is used to estimate attitude, as angular rates (ω = [ωX ωY ωZ ]T ) measured from a gyroscope
shown in Fig. 1. In the prediction step, the EKF estimates attitude fixed on the body frame as follows:
([Φ− , θ− , Ψ− ]) by using the angular rate (ω) measured from the ⎡ ⎤ ⎡ ⎤⎡ ⎤
3-axis gyroscope through a system dynamic model. In the cor- q̇0 q0 −q1 −q2 −q3 0
rection step, the EKF updates roll (Φ+ ) and pitch (θ+ ) by using ⎢ q̇ ⎥ ⎢q ⎥⎢ ⎥
⎢ 1⎥ ⎢ 1 q0 −q3 q2 ⎥ ⎢ ωX ⎥
the acceleration (A) measured from the 3-axis accelerometer q̇ = ⎢ ⎥ = 0.5 ⎢ ⎥⎢ ⎥ . (2)
⎣ q̇2 ⎦ ⎣ q2 q3 q0 −q1 ⎦ ⎣ ωY ⎦
that provides the direction of gravitational acceleration, through
an acceleration measurement model, and updates yaw (Ψ+ ) us- q̇3 q3 −q2 q1 q0 ωZ
ing the magnetic flux density (B) measured from the 3-axis The system dynamic model is discretized with a backward
magnetometer through a magnetometer measurement model. differential to implement the use of digital sensors and micro-
processors as
A. System Dynamic Model A partial derivative of the nonlinear system dynamic model
The system dynamic model is used to predict attitude of [see (3)–(6) shown at the bottom of this page] gives the linearized
a fixed body frame on a mobile platform, with respect to a state transition matrix (Jacobian) to be used for predicting and
navigation frame defined as north, east, and down (N, E, and D). updating covariance matrixes in (7), shown at the bottom of this
Although Euler angles [roll (Φ), pitch (θ), yaw (Ψ)] are more page.

q0,k = q0,k −1 + 0.5 (−q1,k −1 ωX,k −1 − q2,k −1 ωY,k −1 − q3,k −1 ωZ,k −1 ) dt (3)
q1,k = q1,k −1 + 0.5 (q0,k −1 ωX,k −1 − q3,k −1 ωY,k −1 + q2,k −1 ωZ,k −1 ) dt (4)
q2,k = q2,k −1 + 0.5 (q3,k −1 ωX,k −1 + q0,k −1 ωY,k −1 − q1,k −1 ωZ,k −1 ) dt (5)
q3,k = q3,k −1 + 0.5 (−q2,k −1 ωX,k −1 + q1,k −1 ωY,k −1 + q0,k −1 ωZ,k −1 ) dt. (6)
⎡ ⎤
1 −0.5ωX,k −1 dt −0.5ωY,k −1 dt −0.5ωZ,k −1 dt
⎢ 0.5ωX,k −1 dt 1 0.5ωZ,k −1 dt −0.5ωY,k −1 dt ⎥
[1] ⎢
Φk −1 = ⎣ ⎥ (7)
0.5ωY,k −1 dt −0.5ωZ,k −1 dt 1 0.5ωX,k −1 dt ⎦
0.5ωZ,k −1 dt 0.5ωY,k −1 dt −0.5ωX,k −1 dt 1
1410 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 2, FEBRUARY 2018

 
B. Accelerometer Measurement Model [1]
[1] Ha,k
Hk = [1] denotes the combined measurement matrix.
Measurement of accelerations in a body frame can indicate Hm ,k
the direction of gravity which can be used to correct predicted + + + +
Finally, the corrected quaternion (q0, k , q1, k , q2, k , q3, k )
roll (Φ− ) and pitch (θ− ) through the discretized accelerometer can be converted to Euler angles using the following equations:
measurement model (Ha,k ) [16], [29] shown as ⎛  ⎞
⎡ ⎤ +
2 q0,k +
q1,k +
+ q2,k +
q3,k
−2g (q1,k q3,k − q0,k q2,k )
⎢ ⎥ φ = tan−1 ⎝  ⎠ (15)
−2g (q0,k q1,k + q2,k q3,k ) ⎥ + + + +
Ha,k = ⎢⎣  ⎦ (8) 1 − 2 q1,k q1,k + q2,k q2,k
2 2 2
−g q0,k − q1,k − q2,k + q3,k 2  
+ + + +
θ = sin−1 −2 q1,k q3,k − q0,k q2,k (16)
where g is the gravitational acceleration (9.857 m/s2 ). The in- ⎛  ⎞
+ + + +
ertial acceleration terms in (8) are ignored because we use the 2 q1,k q2,k + q0,k q3,k
−1 ⎝  ⎠ .
measured 3-axis acceleration to calculate the direction of grav- ψ = tan (17)
+ + + +
ity only when no inertial acceleration is present (only when 1 − 2 q2,k q2,k + q3,k q3,k
the magnitude of the measured 3-axis acceleration is around
9.857 m/s2 ). The algorithm skips updating roll and pitch angles III. ALGORITHM FOR REJECTING DISTURBED
to avoid errors in the direction of gravity when an inertial accel- MAGNETIC FIELD
eration (magnitude of measured accelerations > 9.857 m/s2 ) is
In this section, we introduce major concepts and methods
detected. Estimation of roll and pitch angles is used for coor-
of the algorithm to detect and reject magnetic disturbances.
dinate transformation from a body frame to navigation frame
More details with illustrations and experimental results are pre-
to compensate effects by terrain roughness when mobile robots
sented in the next section for cases where a particular parame-
move on hills and dales. The linearized Jacobian matrix of the
ter plays an important role in detecting and rejecting magnetic
measurement model is given by
⎡ ⎤ disturbances.
2gq2,k −2gq3,k 2gq0,k −2gq1,k
[1]
Ha, k = ⎣ −2gq1,k −2gq0,k −2gq3,k −2gq2,k ⎦ . A. Magnetic Disturbance
−2gq0,k 2gq1,k 2gq2,k −2gq3,k
(9) An AHRS mounted on a mobile robot updates the heading
angle (Ψ− , yaw) in the correction step with the measured ge-
C. Magnetometer Measurement Model omagnetic north (ψmag ) using the magnetometer measurement
model as shown in Fig. 2(a). However, if the geomagnetic field
Magnetic flux densities measured from a 3-axis magnetome- is interfered with by ferromagnetic objects such as iron or steel,
ter determine geomagnetic north using the following equation 
geomagnetic north (ψmag = ψmag ) changes although the phys-
[16], [29]: ical heading angle is stationary as illustrated in Fig. 2(b). For


−1 −BY example, a stationary mobile robot with a true heading angle of
ψmag = tan (10) 32° reads a magnetic heading angle of 52° when a ferromag-
BX
⎛ ⎞ netic object approaches as shown in Fig. 2(c). The heading angle
2 (q q
1,k 2,k + q q )
0,k 3,k ⎠
(ψgyro ) predicted with angle rates from the gyroscope is not af-
Hm ,k = tan−1 ⎝  . (11) fected by the ferromagnetic objects but drifting occurs because
2
1 − 2 q2,k + q3,k 2
of the integration of errors as shown in Fig. 2(d). The result-
The linearized Jacobian matrix of the magnetometer mea- ing heading angle (ψAHRS ) from the AHRS is also disturbed
surement model is given by as shown in Fig. 2(e) because of the correction steps by the
  disturbed magnetic heading angle (ψmag ) shown in Fig. 2(c).
[1]
Hm ,k = q3,k t0 q2,k t0 q1,k t0 + 2q2,k t1 q0,k t0 + 2q3,k t1 Detecting magnetic disturbances in a geomagnetic field and
(12) skipping correction steps with the disturbed magnetic heading
where t0 and t1 are calculated with the following equations: angle can reduce inaccuracy in the estimated heading angle from
  2 2
 the AHRS. Magnets, ferromagnetic objects on board also distort
t0 = 1 − 2 q2,k + q3,k
geomagnetic flux density such as a constant magnetic field off-
2 set (hard iron distortion) or variable magnetic field offset with
×   2
(2 (q1,k q2,k + q0,k q3,k ))2 + 1 − 2 q2,k
2 + q2 respect to an attitude (soft iron distortion). We compensated
3,k
these offsets by calibrating magnetometers once initially after
(13) implementing the sensors and mobile platforms [20], [21].
t1 = 2 (q1,k q2,k + q0,k q3,k )
2 B. Detection of Magnetic Disturbance
×   2 .
(2 (q1,k q2,k + q0,k q3,k ))2 + 1 − 2 q2,k
2 + q2 To detect a disturbance in the geomagnetic field, we monitor
3,k
three parameters: 1) dip angle, 2) strength of the magnetic flux
(14) density, and 3) difference of the angular rates measured by the
LEE et al.: COMPENSATED HEADING ANGLES FOR OUTDOOR MOBILE ROBOTS IN MAGNETICALLY DISTURBED ENVIRONMENT 1411

Fig. 2. Heading angle estimation with attitude and heading reference system (AHRS) fixed on a mobile robot. (a), (b) Heading angle measurements

with magnetometer in (a) uniform geomagnetic field (heading angle: ψ mag ) and (b) disturbed field (heading angle: ψ mag ). Although the mobile robot

does not change the heading angle physically, the heading angle (ψ mag ) measurement in the disturbed field is different from ψ mag . (c) ψ mag when
a ferromagnetic object approaches the mobile robot at 340 s. (d) Heading angle (ψ gyro ) by integrating angular rates measured from a gyroscope.
(e) Heading angle (ψ AHRS ) estimation corrected by ψ mag .

Fig. 3. Changes of parameters when a magnetic disturbance is present. Measurements of magnetic flux density (BX , BY , BZ ) (a) without and
(b) with magnetic disturbances. ψ mag is the heading angle on the horizontal plane. Although the mobile robot does not change heading, the

magnetometer reads the disturbed heading angle (ψ mag ). The dip angle (ϕ dip ) or the strength of the magnetic flux density (|B|) can be used to
detect a magnetic disturbance. The dip angle and strength of the magnetic flux density should remain constant without magnetic disturbance. In
cases where no changes are detected in the dip angle or strength of the magnetic flux density with magnetic disturbance, the difference between the
angular rate measured by the magnetometer and gyroscope (|Δψ mag − Δψ gyro |) can be used to detect a magnetic disturbance because angular
rate measurement with a gyroscope is not affected by the disturbance.

gyroscope and magnetometer, as illustrated in Fig. 3. A dip angle to detect the magnetic disturbance because angle rate measured
(ϕdip ) is the inclination of the geomagnetic field with respect with a gyroscope is not affected by magnetic disturbance.
to the Earth’s surface as noted in Fig. 3. The dip angle depends
on location, e.g., 51.45° in Gwangju City, South Korea. The
strength of the geomagnetic flux density (|B|) also depends C. Rejection of Disturbed Magnetic Field
on location. The dip angle and strength of the magnetic flux Fig. 4 presents a flow chart of the algorithm that rejects
density should remain constant when no magnetic disturbance a disturbed magnetic heading angle. The algorithm reads the
is present. In cases where no change is detected in the dip angle magnetic flux density and monitors three parameters: dip angle
or the strength of the magnetic flux density with a magnetic (ϕdip ), strength of magnetic flux density (|B|), and difference
disturbance, the difference in angular rates measured by the of angle rates (|Δψmag − Δψgyro |), to see if there are any ab-
gyroscope and magnetometer (|Δψmag − Δψgyro |) can be used normal changes caused by a magnetic disturbance. If all of these
1412 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 2, FEBRUARY 2018

Fig. 4. Block diagram of the algorithm that rejects a magnetic distur-


bance. The algorithm reads the magnetic flux density and checks three
parameters: dip angle (ϕ dip ), strength of magnetic flux density (|B|), and
difference of angle rates (|Δψ mag − Δψ gyro |).

Fig. 5. Thresholds of (a) dip angle, (b) normalized strength of magnetic


parameters are within their thresholds, the magnetic flux density flux density, and (c) difference of angle rates, which were measured with
is considered as not including a magnetic disturbance. There- the magnetometer while moving on a straight path where no magnetic
fore, the magnetic heading angle (ψmag ) is used in the correction disturbances were present.
of the predicted heading angle in the EKF. In other cases, the
algorithm rejects the magnetic flux density and reads a new
magnetic flux density.

D. Thresholds
The thresholds (γ1 , γ2 , γ3 ) shown in Fig. 5 were determined
experimentally. At first, we measured geomagnetic flux den-
sities while moving the magnetometer mounted on a mobile
robot following on a straight path where no magnetic distur-
bances were present. Second, we calculated dip angles (ϕdip ),
normalized strengths (|B|), and differences of angular rates
(|Δψmag − Δψgyro |) with the measured data. Then, we deter-
mined the thresholds, smallest possible, to cover most of the dip
angles, normalized strengths, and differences of angle rates as
shown in Fig. 5. Finally, we verified and adjusted the thresholds Fig. 6. Illustration of a case where magnetic disturbance can be de-
on various places such as on smooth floors, lawn, parking lot, tected by the dip angle. B (red arrow) and B  (green arrow) are the
magnetic flux density with and without disturbance, respectively. When
and sidewalk. Although the thresholds were chosen in the envi- the magnetic flux density changes from B to B  on the surface of the
ronment without magnetic disturbances, the thresholds should sphere, the dip angle changes (ϕ dip = ϕ dip ) but the strength remains
be evaluated and adjusted in other regions if necessary. The constant (|B  | = |B|). Thus, the dip angle can detect that there is a

thresholds we used throughout this paper are γ1 = ±4◦ for magnetic disturbance in the heading angle (ψ mag ).
dip angles, γ2 = ±0.1 for normalized strength, γ3 = 1◦ for
difference of angular rates. former cases can detect a magnetic disturbance more clearly
than the latter cases. In this section, we present how the algo-
IV. RESULTS rithm works in cases where only one parameter out of three (dip
angle, strength, and difference of angle rates) is affected by a
A magnetic disturbance can cause diverse changes in the magnetic disturbance manipulated with permanent magnets.
magnetic flux density, depending on, for example, how ferro-
magnetic objects approach the magnetometer. In many cases,
A. Dip Angle
all three parameters (dip angle, strength of magnetic flux den-
sity, and difference of angle rates) are affected together, or Fig. 6 illustrates a case where a magnetic disturbance causes

in some cases only one or two parameters are affected. The a change in the magnetic heading angle (ψmag = ψmag ) and
LEE et al.: COMPENSATED HEADING ANGLES FOR OUTDOOR MOBILE ROBOTS IN MAGNETICALLY DISTURBED ENVIRONMENT 1413

Fig. 8. Illustration of a case where a magnetic disturbance can be


detected by the strength of magnetic flux density. When magnetic flux
density changes from B to B  on the surface of the cone, the dip angle
remains constant (ϕ dip = ϕ dip ) but the strength changes (|B| = |B|).
Thus, the strength of the magnetic field can detect that there is a mag-
netic disturbance in the heading angle (ψ mag ).

AHRS, the magnetic disturbance is detected by the dip angle as


shown in Fig. 7(a). The dip angle is beyond the threshold (±4◦ ,
dash-dot line) between 1 and 20 s although the normalized
strength (|B|) of the magnetic flux density shown in Fig. 7(b)
and the difference of the angle rates (|Δψmag − Δψgyro |) mea-
sured by a magnetometer and gyroscope shown in Fig. 7(c)
remain mostly within the thresholds. Fig. 7(d) shows the esti-
mated heading angles with (solid line) and without (dotted line)
rejection of the magnetic disturbance. The true heading angle is
zero (dashed line). Without rejecting the magnetic disturbance,
the error in the heading angle goes up to 60°. However, by re-
jecting the magnetic disturbance, the error in the heading angle
is reduced down to smaller than 3.6°.

B. Magnetic Field Strength


Fig. 7. Dip angle, normalized strength of magnetic flux density, differ-
ence of angle rates, and heading angles in a case where the dip angle Fig. 8 shows a case where a magnetic disturbance can be de-
can detect a magnetic disturbance. The AHRS is pointing to geomag- tected by changes in the strength of the magnetic flux density.
netic north without movement. (a) Dip angle when permanent magnets
approach the AHRS. The dip angle is beyond the threshold (dash-dot
When a magnetic disturbance causes changes in the magnetic
line) between 1 and 20 s. (b) Normalized strength (|B|) of the magnetic flux density from B (red solid arrow) to B (green solid ar-
flux density and (c) the difference of the angle rate (|Δψ mag − Δψ gyro |) row) on the surface of the cone, the dip angle remains constant
mostly remain within the thresholds. (d) Estimation of heading angles
with (solid line) and without (dotted line) rejection of the magnetic
(ϕdip = ϕdip ), but the strength of the magnetic flux density
disturbance. changes (|B| = |B|). In this case, the changes in strength of
the magnetic flux density can detect the magnetic disturbance

in the heading angle (ψmag ).
dip angle (ϕdip = ϕdip ) when a vector of the magnetic flux Fig. 9 shows experimental results estimating the heading an-
density (B, red solid arrow) moves to a new vector (B , green gle when a magnetic disturbance is detected by the strength of
solid arrow) because of a magnetic disturbance. B stays on the magnetic flux density. The stationary AHRS is aligned to
the surface of the sphere where B is. In this case, although the point geomagnetic north when permanent magnets approach the
strength of the magnetic flux density remains constant (|B| = AHRS between 1 and 5 s. Although the dip angle changes a rela-
|B|) with the magnetic disturbance, a change in the dip angles tively small amount [see Fig. 9(a)], the magnetic disturbance can
can be used to detect the magnetic disturbance in the magnetic be detected by normalized strength (|B|) as shown in Fig. 9(b).

heading angle (ψmag ). The normalized strength goes up to 3.9°. The difference of the
Fig. 7 shows experimental results reducing estimation errors angle rates (|Δψmag − Δψgyro |) presented in Fig. 9(c) shows
in the heading angle by detecting a magnetic disturbance with peaks when the magnets move to and away from the AHRS.
the dip angle. Physically, the AHRS is stationary and point- Fig. 9(d) shows the estimated heading angles without (dotted
ing geomagnetic north. When permanent magnets approach the line) and with (solid line) rejection of a magnetic disturbance.
1414 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 2, FEBRUARY 2018

Fig. 10. Illustration of a case that a magnetic disturbance cannot be


detected by the dip angle or strength of magnetic flux density. The special
case occurs when the magnetic flux density stays on the edge of the
cone. Although the heading angle is disturbed (ψ mag  = ψ mag ), the dip
angle (ϕ dip = ϕ dip ) and strength of magnetic flux density (|B| = |B|)
stays unchanged.

Fig. 9. Case where the strength of magnetic flux density can detect
a magnatic disturbance. (a) Dip angle when permanent magnets ap-
proach the AHRS between 1 and 5 s. The dip angle stays mostly
within the thresholds. However, (b) normalized strength of magnetic
flux density is beyond the thresholds. (c) The difference of angle rates
(|Δψ mag − Δψ gyro |) changes only slightly when the magnets stay close.
(d) Estimation of heading angles with (solid line) and without (dotted line)
rejecting the magnetic disturbance.

The heading angle estimated by rejecting the magnetic distur-


bance that is detected by the magnetic strength is very close to
the true heading angle (dashed line).

C. Difference of Angular Rates Measured by a


Magnetometer and a Gyroscope
When a magnetic disturbance cannot be detected by the dip
angle or strength of the magnetic flux density when the dis- Fig. 11. Case that a magnetic disturbance cannot be detected by the
dip angle or strength of magnetic flux density. Both of (a) the dip angle
turbed magnetic flux density moves to another point on the edge and (b) normalized strength of magnetic flux density stay within the
of the cone as illustrated in Fig. 10, although the heading an- thresholds. However, (c) difference of angle rates |Δψ mag − Δψ gyro | is

gle is disturbed (ψmag = ψmag ), the dip angle (ϕdip = ϕdip ) out of the throushold. (d) Estimation of heading angles with and without
and strength of the magnetic flux density (|B| = |B|) stay un- rejecting the magnetic disturbance.

changed. The experimental results presented in Fig. 11(a) and


(b) show that the dip angle and normalized |B| stay within the gyroscope (|Δψmag − Δψgyro |) because magnetic disturbances
thresholds even though there are magnetic disturbances induced do not affect the gyroscope. Fig. 11(d) shows the estimated
by approaching permanent magnets near the AHRS. In this heading angles. By rejecting the magnetic disturbances detected
special case, the magnetic disturbance can be detected by the by the difference of the angle rates, the maximum error in the
difference of angular rates measured by the magnetometer and estimated heading angle is lowered almost by half (16◦ → 8◦ ).
LEE et al.: COMPENSATED HEADING ANGLES FOR OUTDOOR MOBILE ROBOTS IN MAGNETICALLY DISTURBED ENVIRONMENT 1415

Fig. 12. Snapshot of the mobile robot for outdoor experiments. The
robot can be controlled manually with a remote radio transmitter or con-
trolled autonomously by detecting obstacles with the light detection and
ranging (LiDAR, UTM-30LX, HUKUYO). The AHRS and control system
are fixed on the center of a mobile robot.

Fig. 14. Experimental results on lawn with apparent magnetic distur-


bances. (a) Four ferromagnetic objects placed every 2 m on a straight
path provide apparent magnetic disturbances. (b) Four peaks of the dip
angle beyond the thresholds indicate magnetic disturbances. (c) Nor-
malized strength of magnetic flux density. (d) Difference of angle rates
(|Δψ mag − Δψ gyro |). Higher peaks are observed near the ferromag-
netic objects. (e) Heading angles calculated without (dotted line) and
with (solid line) rejection of the magnetic disturbances.

D. Outdoor Experiments on a Lawn


The performance of the algorithm is also evaluated while driv-
ing a mobile robot on a lawn with and without magnetic distur-
bances. The AHRS with the embedded algorithm is mounted on
the center of a mobile robot (size ∼ 34 cm × 36 cm × 24 cm)
as shown in Fig. 12. We used two channels of a radio transmitter
to control the throttle and steering. Duty ratio of radio signals
determines the wheel speed. Throughout the experiments, the
Fig. 13. Experimental results on lawn without apparent magnetic dis-
turbances. (a) A straight path on a lawn is used to evaluate performance mobile robot was manually controlled by a pilot. On a lawn, the
of the algorithm mounted on a mobile robot. (b) Dip angle during a mobile robot is driven on a straight path (10 m) at a speed of
straight run. The threshold (dash-dot lines, 51.45 ± 4◦ ) is determined 0.57 m/s.
to cover the variation (∼5°) caused by the uneven surface of the lawn.
(c) Normalized strength of magnetic flux density |B|. The normalized |B| Fig. 13(a) shows the mobile robot running on a lawn with-
is mostly within the threshold (dash-dot lines 1 ± 0.1). (d) Difference of out any apparent ferromagnetic objects in the straight path. The
angle rates (|Δψ mag − Δψ gyro |) measured by the magnetometer and dip angle (ϕdip ), normalized strength of magnetic flux den-
gyroscope. The threshold (1°, dash-dot line) covers most of the differ-
ence. (e) Heading angles calculated without (dotted line) and with (solid sity (|B|), and difference of angle rates (|Δψmag − Δψgyro |)
line) rejection of the magnetic disturbance. are mostly within each threshold (dash-dot line) as shown in
1416 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 2, FEBRUARY 2018

Fig. 15. Experimental results of the algorithm mounted on a mobile robot running through (a) a square path (10 m × 10 m) on a smooth
ferroconcrete rooftop by taking 10 turns. (b) Dip angle, (c) normalized strength of magnetic flux density, and (d) difference of angle rates measured
by the magnetometer and gyroscope. Repetitive measurements show similar trends in the same locations, indicating that the variations out of the
thresholds are caused by ferromagnetic objects in the concrete or under the rooftop. (e) True heading angles (dashed line), and estimated heading
angles (dotted line) without and (solid line) with rejection of magnetic disturbance. The dash-dotted line denotes the estimated heading angle with
a gyroscope only.

TABLE I
PEAK-TO-PEAK AND RMS ERROR OF THE HEADING ANGLE ON THE FLAT ROOF OF A FERROMAGNETIC BUILDING

1st turn 2nd turn 3rd turn 4th turn 10th turn
Time (0–48 s) (48–92 s) (92–138 s) (138–184 s) (413–465 s)

Peak-to-Peak Without Algorithm (dashed line) 61.0° 59.2° 62.1° 56.0° 57.8°
With Algorithm (solid line) 20.2° 18.7° 22.5° 22.4° 28.1°
RMS Without Algorithm (dashed line) 13° 8.0° 7.8° 6.1° 4.7°
With Algorithm (solid line) 2.8° 2.0° 1.8° 1.8° 1.5°

Fig. 13(b)–(e). The algorithm considers there to be no signifi- cause large deviation in the heading angle as shown with the dot-
cant magnetic disturbance to reject in the path. As a result, the ted line in Fig. 14(e). For example, the deviation in the heading
heading angles estimated without and with rejection of magnetic angle around 4 s is about 50°, which is much higher than the
disturbances are similar in Fig. 13. The deviations (< ±7◦ ) of deviation (< ±7◦ ) caused by an uneven surface on the field.
the heading angles are caused by real changes of the heading of The difference in the deviations of the heading angle by each
the mobile robot due to the uneven surface of the field. Fig. 14(a) ferromagnetic object is attributed to the different orientation
shows the mobile robot running over square ferromagnetic ob- of magnetic disturbances because geometrically identical ferro-
jects (size: 150 mm × 150 mm × 10 mm, steel plates) placed magnetic objects with the same composition can have randomly
every 2 m in a straight path. The algorithm clearly detects the oriented magnetic domains on the ground since the geomag-
magnetic disturbances with the dip angle [ϕdip , Fig. 14(b)], nor- netic field is too weak to uniformly magnetize ferromagnetic
malized strength [|B|, Fig. 14(c)], and difference of angular rates objects [30]. By rejecting the magnetic disturbances, the algo-
[|Δψmag − Δψgyro |, Fig. 14(d)]. The magnetic disturbances can rithm reduces the deviation [< ±7◦ , Fig. 14(e)] similar to the
LEE et al.: COMPENSATED HEADING ANGLES FOR OUTDOOR MOBILE ROBOTS IN MAGNETICALLY DISTURBED ENVIRONMENT 1417

deviation [< ±7◦ , Fig. 13(e)] measured without disturbances


by the apparent ferromagnetic objects.

E. Outdoor Experiments on a Flat Roof


In the other experiments, the mobile robot is manually con-
trolled to follow a square path (10 m × 10 m) in a clockwise
manner for ten turns on a flat smooth ferroconcrete roof that
contains iron bars. The surface of the roof coated with thick
waterproof paint is relatively very smooth. The iron bars on
the roof or other ferromagnetic objects and electric wires un-
der the roof can cause magnetic disturbances as detected with
the dip angle (ϕdip ), normalized strength (|B|), and difference
of angular rates (|Δψmag − Δψgyro |) in Fig. 15(b)–(d). These
magnetic disturbances cause errors (up to 22.2°) of the heading
angle as plotted with the dotted line in Fig. 15(e). The experi-
mental results indicate that by rejecting the detected magnetic
disturbances (solid line), the algorithm improves the accuracy of
the heading angle and brings it much closer to the true heading
angle as given with the peak-to-peak and rms errors in Table I.
Although the algorithm skips correction steps [noted with dots
in Fig. 15(e)] continuously for several seconds due to magnetic
disturbances, the intermittent correction steps in EKF prevent
the heading angle from drifting, in contrast to the heading an-
gle estimated only with gyroscope (17.4° of drift for 460 s) as
shown with the dash-dotted line in Fig. 15(e).

F. Outdoor Experiment on a Parking Lot


The performance of the method is also evaluated under more
realistic environment such as in a parking lot as shown in Fig. 16.
The mobile robot is manually controlled to move down (in-
clination: ∼20°) from the point on a lawn to the parking lot,
and to pass by the cars whose ferromagnetic obstacles induce
magnetic disturbances. Then, the mobile robot moves back to
the starting point on a lawn at average speed of ∼1.3 m/s as
shown in Fig. 16(a). This scenario simulates realistic use of a
mobile robot to deliver packages from a parking lot to a front Fig. 16. Experimental results in the parking lot while driving the mobile
yard. The magnetic disturbances caused by the ferromagnetic robot through (a) a path (red line) at an average speed of 1.3 m/s.
obstacles in the parking lot, similarly caused by ferromagnetic The mobile robot started from the point on a lawn, moved down to the
parking lot and passed by the cars, which induce magnetic disturbances.
materials underground, induce the fluctuations in the dip an- Then, the mobile robot was driven to go back to the starting point.
gle (ϕdip ), normalized strength (|B|), and difference of angular (b) Dip angle, (c) normalized strength of magnetic flux density (|B|), and
rates (|Δψmag − Δψgyro |) shown in Fig. 16(b)–(d). The algo- (d) difference of angle rates (|Δψ mag − Δψ gyro |). (e) Estimated heading
angles with (solid line) and without (dotted line) rejection of the magnetic
rithm by detecting and rejecting magnetic disturbances reduces disturbances.
the deviations of the heading angles (solid line, peak-to-peak:
10.7°, rms: 2.3°), compared to the results without the algo-
rithm (dotted line, peak-to peak: 29.9°, rms: 5.5°) as shown in ∼ 34 cm × 36 cm × 24 cm) is implemented with a light
Fig. 16(e). The algorithm increased the computation time and detection and ranging sensor to detect and avoid obstacles as
CPU usage by 3.2 µs and 5.7%, respectively, when measured shown in Fig. 12. The heading angle estimated by the proposed
with the microprocessor (32 bit STM32F427 Cortex M4 core, algorithm is used to control wheel speeds of the mobile robot
168 MHz, 256 Kbyte RAM). for autonomous driving on the sidewalk where a manhole cover,
bushes, cones, bike, motorbike, and a man walking present
static, dynamic obstacles or magnetic disturbances. Initially, the
G. Outdoor Experiment with an Autonomous Robot with
heading angle of the mobile robot is –105° and the desired head-
Obstacle Avoidance ing angle is changing to –55°, –155°, –35°, and –175° sequen-
The algorithm is also evaluated with an autonomous tially as shown in the region 1 in Fig. 17(a). The autonomous
robot in more realistic scenarios on the sidewalk of the robot follows the desired angles (dashed line) sequentially as
road as shown in Fig. 17. The autonomous robot (size shown in Fig. 17(e). In regions 2 ,3 , and 
5 , a manhole cover,
1418 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 65, NO. 2, FEBRUARY 2018

TABLE II
PEAK-TO-PEAK AND RMS ERROR OF THE HEADING ANGLE ON
MOTORIZED STAGE

Yadav Roetenberg Madgwick Proposed


et al. et al.

Uncompensated Peak-to-Peak 23° 23° 23° 129.9°


RMS 15.4° 15.4° 15.4° 14.4°
Compensated Peak-to-Peak 5° 12° 10° 6.6°
RMS 3.4° 9.3° 6° 2.0°

It should be noted that our experiments provide a rough com-


parison since we do not have the other algorithms and systems
required for a more refined side-by-side comparison. We started
the experiments by fixing and calibrating the magnetometer to
the center of the motorized rotation stage. For a clean start,
we initially calibrated the magnetometer to reduce unnecessary
magnetic disturbances from the motorized stage that includes
an electric motor, steel plates, and poles [20]. While intention-
ally adding magnetic disturbances, by approaching a permanent
magnet, to make the uncompensated rms error similar to others
(∼15°), we rotated the magnetometer on the motorized stage
by ±90° at a speed of 5°/s on a level surface for 320 s. The
high peak-to-peak error (129.9°) of the uncompensated case is
reduced down to 6.6° by applying the proposed algorithm. The
rms results also show that the algorithm we proposed provides
similar or slightly better accuracy as in Table II.

V. CONCLUSION
The results reported here showed that an algorithm rejecting
magnetic disturbances improved accuracy of heading angles es-
timated with the EKF. The dip angle, strength of the magnetic
flux density, and difference of the angular rates measured by
the magnetometer and gyroscope were complementary to each
other. For cases where one parameter cannot be detected, other
parameters detected magnetic disturbances. Experiments using
the algorithm implemented in a relatively low-cost micropro-
Fig. 17. Experimental results with an autonomous robot that detects
obstacles and avoids collision (a) on the sidewalk (red line). The wheel cessor and sensor module on a lawn and flat roof demonstrated
speeds of the mobile robot are controlled by using the desired and esti- the feasibility of the methods by comparing the estimation re-
mated heading angles. (b) Dip angle, (c) normalized strength of magnetic sults in EKF on mobile robots with and without implementation
flux density (|B|), and (d) difference of angle rates (|Δψ mag − Δψ gyro |).
(e) Estimated heading angles with (solid line) and without (dotted line) re- of the algorithm. These results should be useful for diverse
jecting magnetic disturbances. The dashed line notes the desired head- applications using mobile robots, autonomous vehicles, and air-
ing angle. craft, which are becoming more popular and should be operating
in magnetically disturbed environment such as in metropolitans
motorbike, and bike cause magnetic disturbances which are with lots of cars, buildings, and electrical cables for autonomous
detected by the dip angle, normalized strength, or difference of transportation, package delivery, surveillance, and many others.
angular rates [see Fig. 17(b)–(d)] and rejected by the algorithm The proposed method improves accuracy of the heading an-
[see Fig. 17(e)] while the mobile robot steers autonomously gle without additional sensors, thus, should be beneficial in
to avoid static (4 cones) and dynamic ( 3 motorbike,  5 bike, enhancing performances of mobile robots’ essential function-
6 pedestrian) obstacles. alities such as localization, real-time obstacle avoidance, and
target tracking [31], [32].
H. Comparison
The proposed algorithm is evaluated indoors on a motorized REFERENCES
stage by changing the heading angle by 90° to compare with [1] R. R. Murphy, “Human-robot interaction in rescue robotics,” IEEE Trans.
the previous results [22], [24], [25] as in a previous report [25]. Syst. Man Cybern. C, Appl. Rev., vol. 34, no. 2, pp. 138–153, May 2004.
LEE et al.: COMPENSATED HEADING ANGLES FOR OUTDOOR MOBILE ROBOTS IN MAGNETICALLY DISTURBED ENVIRONMENT 1419

[2] M. Schwager, B. J. Julian, M. Angermann, and D. Rus, “Eyes in the sky: [25] N. Yadav and C. Bleakley, “Accurate orientation estimation using AHRS
Decentralized control for the deployment of robotic camera networks,” under conditions of magnetic distortion,” Sensors, vol. 14, pp. 20008–
Proc. IEEE, vol. 99, no. 9, pp. 1541–1561, Sep. 2011. 20024, Oct. 2014.
[3] Zhang and J. M. Kovacs, “The application of small unmanned aerial [26] M. Park and Y. Gao, “Error analysis and stochastic modeling of low-cost
systems for precision agriculture: A review,” Precis. Agric., vol. 13, no. 6, MEMS accelerometer,” J. Intell. Robot. Syst., vol. 46, no. 1, pp. 27–41,
pp. 693–712, Jul. 2012. May 2006.
[4] J. Fink, V. Bauwens, F. Kaplan, and P. Dillenbourg, “Living with a vacuum [27] D. O. Anderson and J. B. Moore, Optimal Filtering. North Chelmsford,
cleaning robot: A 6-month ethnographic study,” Int. J. Soc. Robot., vol. 5, MA, USA: Courier, 2012, pp. 105–106, 194–195.
no. 3, pp. 389–408, Jun. 2013. [28] J. Diebel, “Representing attitude: Euler angles, unit quaternions, and ro-
[5] C. Luo, S. X. Yang, X. Li, and M. Q. Meng, “Neural-dynamics-driven tation vectors,” Stanford Univ., Stanford, CA, USA, Tech. Rep., 2006.
complete area coverage navigation through cooperation of multiple mobile [29] J. B. Kuipers, Quaternions and Rotation Sequences, vol. 66. Princeton,
robots,” IEEE Trans. Ind. Electron., vol. 64, no. 1, pp. 750–760, Jan. 2017. NJ, USA: Princeton Univ., 1999, pp. 127–143.
[6] A. Amanatiadis, “A multisensor indoor localization system for biped [30] H. J. Williams, R. M. Bozorth, and W. Shockley, “Magnetic domain
robots operating in industrial environments,” IEEE Trans. Ind. Electron., patterns on single crystals of silicon iron,” Phys. Rev., vol. 75, no. 1938,
vol. 63, no. 12, pp. 7597–7606, Dec. 2016. pp. 155–178, Jan. 1949.
[7] R. Pfeifer, M. Lungarella, and F. Iida, “Self-organization, embodiment, [31] H. Xiao, Z. Li, and C. L. P. Chen, “Formation control of leader–
and biologically inspired robotics,” Science, vol. 318, no. 5853, pp. 1088– follower mobile robots’ systems using model predictive control based
1093, Nov. 2007. on neural-dynamic optimization,” IEEE Trans. Ind. Electron., vol. 63,
[8] B. Kingston and A. W. Beard, “Real-time attitude and position estimation no. 9, pp. 5752–5762, Sep. 2016.
for small UAVs using low-cost sensors,” in Proc. AIAA Unmanned Unltd. [32] X. Zhang, Y. Fang, and X. Liu, “Motion-estimation-based visual servoing
Tech. Conf. Work. Exhib., Sep. 2004, pp. 2004–6488. of nonholonomic mobile robots,” IEEE Trans. Robot., vol. 64, no. 1,
[9] Y. Kang, C. Roh, S.-B. Suh, and B. Song, “A lidar-based decision-making pp. 390–400, Jan. 2017.
method for road boundary detection using multiple Kalman filters,” IEEE
Trans. Ind. Electron., vol. 59, no. 11, pp. 4360–4368, Nov. 2012.
[10] J. Seron, J. L. Martinez, A. Mandow, A. J. Reina, J. Morales, and A.
J. Garcia-Cerezo, “Automation of the arm-aided climbing maneuver for Jehong Lee (GS’15) received the B.Sc. de-
tracked mobile manipulators,” IEEE Trans. Ind. Electron., vol. 61, no. 7, gree in radio communication engineering from
pp. 3638–3647, Jul. 2014. Korea Maritime and Ocean University, Busan,
[11] K.-B. Lee, H. Myung, and J.-H. Kim, “Online multiobjective evolutionary South Korea, in 2012, and the M.Sc. degree in
approach for navigation of humanoid robots,” IEEE Trans. Ind. Electron., mechatronics from the Gwangju Institute of Sci-
vol. 62, no. 9, pp. 5586–5597, Sep. 2015. ence and Technology, Gwangju, South Korea,
[12] J. Yuan, F. Sun, and Y. Huang, “Trajectory generation and tracking control in 2014, where he is currently working toward
for double-steering tractor–trailer mobile robots with on-axle hitching,” the Ph.D. degree at the School of Mechanical
IEEE Trans. Ind. Electron., vol. 62, no. 12, pp. 7665–7677, Dec. 2015. Engineering.
[13] J. Kim and W. Chung, “Localization of a mobile robot using a laser range His research interests include mobile robots,
finder in a glass-walled environment,” IEEE Trans. Ind. Electron., vol. 63, bio-inspired robots, navigation systems, and
no. 6, pp. 3616–3627, Jun. 2016. sensor fusion.
[14] J. L. Marins, X. Yun, E. R. Bachmann, R. B. Mcghee, and M. J. Zyda,
“An extended Kalman filter for quaternion-based orientation estimation
using MARG sensors,” in Proc. IEEE/RSJ Int. Conf. Intell. Robot. Syst.,
Maui, HI, Oct./Nov. 2001, pp. 2003–2011.
[15] M. L. Psiaki, F. Martel, and P. K. Pal, “Three-axis attitude determination Jeonggeun Lim (GS’16) received the B.Sc.
via Kalman filtering of magnetometer data,” J. Guid. Control Dyn., vol. 13, degree in mechanical design engineering from
no. 3, pp. 506–514, May 1990. Chungnam National University, Daejeon, South
[16] D. Titterton and John L. Weston, Strapdown Inertial Navigation Technol- Korea, in 2012, and the M.Sc. degree in me-
ogy, vol. 17. London, U.K.: IET, 2004, pp. 310–333. chanical engineering from the Gwangju Institute
[17] J. Vaganay, M. J. Aldon, and A. Fournier, “Mobile robot attitude estima- of Science and Technology, Gwangju, South Ko-
tion by fusion of inertial data,” in Proc. IEEE Int. Conf. Robot. Autom., rea, in 2016, where he is currently working to-
Atlanta, GA, May 1993, pp. 277–282. ward the Ph.D. degree at the School of Mechan-
[18] N. Metni, J.-M. Pflimlin, T. Hamel, and P. Souères, “Attitude and gyro ical Engineering.
bias estimation for a VTOL UAV,” Control Eng. Pract., vol. 14, no. 12, His research interests include mobile robots,
pp. 1511–1520, Dec. 2006. unmanned aerial vehicles, and bio-inspired and
[19] M. J. Caruso, “Applications of magnetoresistive sensors in navigation soft robots.
systems,” SAE Tech. Paper, vol. 72, pp. 15–21, 1997.
[20] M. J. Caruso, “Applications of magnetic sensors for low cost compass
systems,” in Proc. IEEE 2000 Position Location Navig. Symp., San Diego,
CA, Mar. 2000, pp. 1–8.
[21] Gebre-Egziabher, G. Elkaim, J. Powell, and B. Parkinson, “A non-linear, Jongho Lee (M’15) received the B.Sc. degree
two-step estimation algorithm for calibrating solid-state strapdown mag- in mechanical engineering from Hanyang Uni-
netometers,” in Proc. 8th Int. Conf. Integr. Navig. Syst., St. Petersburg, versity, Seoul, South Korea, in 1998, the M.Sc.
Russia, May 2001, pp. 290–297. degree in mechanical engineering from Korea
[22] D. Roetenberg, H. J. Luinge, C. T. M. Baten, and P. H. Veltink, “Compen- Advanced Institute of Science & Technology,
sation of magnetic disturbances improves inertial and magnetic sensing Daejeon, South Korea, in 2003, and the M.Sc.
of human body segment orientation,” IEEE Trans. Neural Syst. Rehabil. degree in electrical engineering and computer
Eng., vol. 13, no. 3, pp. 395–405, Sep. 2005. sciences and the Ph.D. degree in mechani-
[23] J. Bird and D. Arden, “Indoor navigation with foot-mounted strapdown in- cal engineering from the University of Califor-
ertial navigation and magnetic sensors,” IEEE Wireless Commun., vol. 18, nia, Berkeley, CA, USA, in 2008 and 2006,
no. 2, pp. 28–35, Apr. 2011. respectively.
[24] S. O. H. Madgwick, A. J. L. Harrison, and R. Vaidyanathan, “Estimation He is currently an Associate Professor with the School of Mechanical
of IMU and MARG orientation using a gradient descent algorithm,” in Engineering, Gwangju Institute of Science and Technology, Gwangju,
Proc. IEEE Int. Conf. Rehabil. Robot., Zurich, Swiss, Jun./Jul. 2011, South Korea. His research interests include mobile robots, bio-inspired
pp. 179–185. and bio-integrated systems.

You might also like