2013 International Conference on Unmanned Aircraft Systems (ICUAS)

May 28-31, 2013, Grand Hyatt Atlanta, Atlanta, GA

Basics and Guidelines of Complementary Filters

for Small UAS Navigation

Austin Jensen, Cal Coopmans YangQuan Chen

Utah Water Research Laboratory Mechatronics, Embedded Systems
Utah State University and Automation (MESA) Lab
Logan, UT 84321 University of California, Merced
Email: [email protected] Merced, CA 95343
Web: http://aggieair.usu.edu Email: [email protected]

Abstract—Orientation estimation is very important for de- High-pass

velopment of unmanned aerial systems (UAS). Kalman filters Sensor A
Filter
are widely used; however they assume linearity and Gaussian
statistics. While these assumptions work well for high-quality, x + x̂
high-cost sensors, they do not work as well for low-cost, low-
quality sensors. In these cases, complementary filters are used
since no assumptions are made with regards to linearity and Low-pass
Sensor B
statistics. This paper gives a review of the different types Filter
of complementary filters being developed. Basic examples are
shown with details about how they are derived and simulations Fig. 1. Simple Complementary Filter Diagram.
that show how they work. Guidelines are then given about when
complementary filters are best used for UAS navigation.
shows a diagram of a simple complementary filter with two
sensors.
I. I NTRODUCTION
Complementary filters were first introduced in a patent
Estimating the orientation of the aircraft is a very im- in 1951 [12]. The patent describes an automated steering
portant part of an unmanned aerial system (UAS). Com- system for keeping a manned aircraft on a proper flight
monly, a Kalman filter is used [1]–[5], which combines path during landing using the Instrument Landing Approach
information from multiple sensors and models to estimate System (ILAS) along with a gyro-stabilized accelerometer.
the orientation and variance. However, when working with The ILAS was used to measure the rate of change between
small, low-quality inertial sensor systems, Kalman filters the distance from the aircraft and the desired path, and the
can be difficult to apply due to the non-linearities and non- gyro-stabilized accelerometer was used to measure changes
Gaussian characteristics of these sensors [6], [7]. In these in heading. The ILAS had reliable low frequency data but
cases, complementary filters are often used [8]–[10] because unreliable high frequency data, while the accelerometer had
no assumptions are made about linearity and noise statistics reliable high frequency data. Therefore, a complementary
[11]. This paper reviews complementary filters and discusses filter was used to combine this data, although the term
how they have formed over the years, along with their “complementary filter” was not used. This term was coined
applications to UAS navigation. in a journal paper in 1953 that describes a similar automated
The basic idea of a complementary filter involves taking steering system [13]. The complementary filters in [12]
two or more sensors, filtering out unreliable frequencies and [13] were both designed in the frequency domain and
for each sensor, and combining the filtered outputs to get realized with simple RC circuits.
a better estimate throughout the entire bandwidth of the Many other applications for the Frequency Domain Com-
system. To achieve this, the sensors included in the filter plementary Filter (FDCF) were presented after their intro-
should complement one another by performing better over duction. Shaw and Srinivasan [14] used a complementary
specific parts of the system bandwidth. For example, if one filter to combine low frequency position data from strain
sensor has reliable data at high frequencies and unreliable sensors and high frequency data from an accelerometer to
data at low frequencies, one or more sensors should be estimate the position of the end of a beam. Zimmermann
added to the system that have reliable low frequency data. and Sulzer [15] used low frequency data from an incli-
An example of two sensors that complement one another nometer and high frequency accelerometer data to measure
are an accelerometer and a rate gyro. While estimating orientation. Baerveldt and Klang [8] introduced a similar
orientation, the gyro is reliable at high frequencies, and system to measure orientation for an autonomous helicopter
the accelerometer is reliable at lower frequencies [6]. The with an inclinometer and a gyro. Roberts et. al. [9] used an
complementary filter eliminates the unreliable frequencies accelerometer and a gyro to measure 3D orientation.
for each sensor and then combines their output. Figure 1
978-1-4799-0817-2/13/$31.00 ©2013 IEEE 500
 xm1 Bode Plot: =1
0
H1 (s) G1 (s)
0

Magnitude (dB)
x + x̂ −50
G1
xm2 G2
H2 (s) G2 (s) −100 −2
10 10
−1
10
0 1
10
2
10
Frequency (rad/s)

200
Fig. 2. Frequency Domain Complementary Filter

Phase (deg)
100

Eventually researchers started to design complementary 0

filters in state space due to their advantages with multiple-

−100 −2
input multiple-output systems, realization in digital systems, 10 10
−1
10
0 1
10
2
10
Frequency (rad/s)
and closed loop control designs. A common application of
the State Space Complementary Filter (SSCF) is to estimate Fig. 3. Bode plots of G1 (s) and G2 (s).
the orientation of a rigid body. Bachmann et. al. [16]–[18]
used such a complementary filter based on quaternions to
track the orientation of robotic flexible links. Quaternions that H1 (s) and H2 (s) are both ideal (H1 (s) = H2 (s) = 1).
are used to represent orientation because they represent the Instead of specifically designing G1 (s) and G2 (s) based on
axis and the rotation angle directly without singularities the sensor models, G1 (s) and G2 (s) are set as generic filters
[19]. Instead of using quaternions to represent orientation, and tuned to minimize the error (e = x − x̂).
Mahony et. al. [6] introduced a direct and passive com-
plementary filter using a rotation matrix in the special With H1 (s) = H2 (s) = 1, Eq. (1) becomes the
orthogonal group, SO(3) (also referred to as a direct cosine following:
matrix (DCM) [20], or attitude matrix [19]). Although the
design of the filter was carried out using the SO(3), Mahony G1 (s) + G2 (s) = 1, ∀s, (2)
et. al. [21] recommended that realization be done with
quaternions. Hamel and Mahony [10] introduced another
and is used as a constraint when choosing filters for
type of complementary filter based on SO(3), called the
G1 (s) and G2 (s).
explicit complementary filter.
While using the modeless approach, one might choose
This study illustrates how different complementary filters the second order filters in Eqs. (3) and (4) for G1 (s) and
are derived. Examples of each type of filter are simulated G2 (s). Since ω0 determines where the filters are cut-off, it
to show how they work and to make comparisons. The is used as a tuning parameter. Figure 3 shows a Bode plot
frequency domain complementary filter is described first in of G1 (s) and G2 (s) with ω0 = 1. If ω0 is reduced, more
Section II. Then examples of state space complementary of H1 (s) is filtered out and more of H2 (s) is allowed to
filters are given in Section III. Section IV presents guidelines pass through (H2 (s) dominant). If ω0 is increased, more of
that describe the best situations to use complementary filters H2 (s) is filtered out and more of H1 (s) is allowed to pass
and conclusions are discussed in Section V. through (H1 (s) dominant). A well-designed filter will have
a value for ω0 such that the best parts of both H1 (s) and
II. F REQUENCY D OMAIN C OMPLEMENTARY F ILTERS H2 (s) are allowed to passed through while the unreliable
parts are filtered out.
To illustrate how a FDCF works, assume that two sensors
complement each other and can be used to measure a signal
x. Both of these sensors can be modeled and have transfer 2ω0 s + ω02
G1 (s) = . (3)
functions H1 (s) and H2 (s). According to the complemen- (s + ω0 )2
tary filter design, the unreliable frequencies of each sensor
are filtered out by passing the outputs through filters G1 (s) s2
and G2 (s). Figure 2 shows a block diagram of this system G2 (s) = . (4)
(s + ω0 )2
[8], [22].
The objective of the estimator is to find an estimate of A. FDCF Example - Modeless Second Order Filter
x (x̂) such that x̂ = x. This is possible, according to the
block diagram in Fig. 2, if the transfer functions G1 (s) and A modeless FDCF with second order filters can be used
G2 (s) are designed according to the following equation: to combine the outputs of an inclinometer and a gyroscope
to measure orientation. An inclinometer is a general term
given to devices that directly measure orientation; one
H1 (s)G1 (s) + H2 (s)G2 (s) = 1, ∀s. (1) example is an accelerometer. The accelerometer measures
orientation by the acceleration due to gravity. In general,
If H1 (s) and H2 (s) are known, they can be used with inclinometers have a slow response time and will filter
Eq. (1) to find G1 (s) and G2 (s). If H1 (s) and H2 (s) are out high frequency movement. A gyro measures angular
not known, a modeless approach can be used which assumes rate, which is converted to orientation by integration. While
501
 2
Inclinometer Output vs. Actual Orientation: =26.985
15
Actual Orientation
Inclinometer Output

10

Orientation (deg)
0

−5

−10

Fig. 4. Simulink Model of Complementary Filter - Modeless Design.

−15
0 50 100 150
Time (s)

the gyro does well at high frequencies, the inherent noise

and integration introduces drift. The complementary filter Fig. 5. Response of the inclinometer to the input chirp signal.
improves the orientation estimate by combining the reliable
high frequency data from the gyro and the low frequency Integrated Gyro Output vs. Actual Orientation: 2
=6805.1822
160
data from the inclinometer. Actual Orientation
Integrated Gyro Output
140
Simulink is used to test the Second Order FDCF with the
inclinometer and gyro. As shown in the simulink model in 120

Fig. 4, the inclinometer is modeled as the transfer function 100

Orientation (deg)
in Eq. (5) [8]. The gyro is modeled as the derivative of the
80
actual orientation plus a bias. Noise is also added to the gyro
before the signal is integrated to get orientation. The output 60

of the sensors are then passed through the filters from Eqs. 40
(3) and (4) and combined to get the orientation estimate.
20

1.88 0
H1 (s) = . (5)
s + 1.88 −20
0 50 100 150
Time (s)

The input orientation to the simulation is a chirp sig-

nal that produces a sine wave between ±10 degrees at Fig. 6. Response of the gyro to the input chirp signal.
frequencies between 0.05Hz and 1Hz. At the beginning
of the simulation, the sine wave will start at 1Hz and Variance vs. Cutoff Frequency
gradually decrease to 0.05Hz after 100 seconds. At that 14
point, the frequency will gradually increase until the end
of the simulation. 12
Gyro-dominant Inclinometer-dominant
Figures 5 and 6 show the response of the inclinometer
10
and gyro to the input chirp signal. As expected, the incli-
nometer cannot track the input at high frequencies. However,
V ariance

8
as the frequency is reduced, the inclinometer is more capable
of tracking the sine wave. The output of the gyro is also
6
expected; it tracks the input well but drifts away from the
actual orientation at 1 degree per second.
4
The filter is tuned by finding the optimal value of ω0 X: 0.22
Y: 2.231
that minimizes the error e. This is done by running multiple 2
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
simulations with varying values of ω0 (0.01 rad/s to 0.5 C utoff F requency (rad/s ec)
rad/s). At the end of each simulation, the squared error and
the variance are calculated for each value of ω0 . Figure 7 Fig. 7. Variance of the error vs. cutoff frequency (ω0 ).
shows the result of these simulations. The lowest variance
is σ 2 = 2.23 with an ω0 of 0.22 rad/s. Figures 8 and 9
show plots of the filter output (estimated orientation) and
the squared error with this optimal frequency. Figures 10 and 11 show the filter output and error plots
when ω0 is too high and the output is dominated by the
While exploring the concept of a complementary filter, inclinometer. Notice how similar the output plot matches
it is also worthwhile to look at the filter output at non- the output of the inclinometer in Fig. 5. Figures 12 and 13
optimal frequencies. As Fig. 7 explains, if ω0 is too high, show the filter output and error plots when ω0 is too low and
the inclinometer will dominate the output of the filter. If ω0 the output is dominated by the gyro. These gyro dominated
is too low, the gyro will dominate the output of the filter. plots also show similarities to the gyro output in Fig. 6.
502
 2
Estimated Output vs. Actual Orientation: =0.22 Estimation Error: =3, =26.7165
0 0
15 120
Actual Orientation
Estimated Output

10 100

5 80
Orientation (deg)

Error Squared
0 60

−5 40

−10 20

−15 0
0 50 100 150 0 50 100 150
Time (s) Time (s)

Fig. 8. Estimated orientation vs. actual orientation at ω0 = 0.22 rad/s Fig. 11. Squared error for ω0 = 3 rad/s (Inclinometer Dominant).
(Optimal Frequency).

2 Estimated Output vs. Actual Orientation: =0.01

Estimation Error: =0.22, =2.2305 0
0
25 60
Actual Orientation
Estimated Output
50
20

40

Orientation (deg)
15
Error Squared

30

20
10

10

5
0

0 −10
0 50 100 150 0 50 100 150
Time (s) Time (s)

Fig. 9. Squared error for ω0 = 0.22 rad/s (Optimal Frequency). Fig. 12. Estimated orientation vs. actual orientation at ω0 = 0.01 rad/s
(Gyro Dominant).
Estimated Output vs. Actual Orientation: 0
=3
15
Actual Orientation
2
Estimated Output Estimation Error: =0.01, =909.2397
0
1800
10

1600

5
1400
Orientation (deg)

1200
0
Error Squared

1000

−5
800

600
−10

400

−15
0 50 100 150 200
Time (s)
0
0 50 100 150
Time (s)
Fig. 10. Estimated orientation vs. actual orientation at ω0 = 3 rad/s
(Inclinometer Dominant).
Fig. 13. Squared error for ω0 = 0.01 rad/s (Gyro Dominant).

B. Other FDCF Design Options

The previous example simulated for this study shows include more complex, higher order filters and filters based
how a modeless FDCF with second order filters can be on the sensor models. Some of these options may achieve
used to combine the output of two sensors to get a better better performance than was shown in the previous example
estimate of orientation. Other design options for FDCFs [8].
503
 Variance vs. Gain
φi 25
Gi (s)

20
+ φ̂

φg 1 15

Variance
s Gi (s)
10

Fig. 14. Frequency Domain Complementary Filter.

X: 0.84
Y: 4.15
5
φg

0
φi + 1
φ̂ 0 0.5 1 1.5 2 2.5 3
+
−
C(s) +
s Gain

Fig. 16. Variance of the error vs. controller gain (P ) for SSCF.

Fig. 15. Closed-loop Complementary Filter System. Estimated Output vs. Actual Orientation: P=0.84
15
Actual Orientation
Estimated Output

10
III. S TATE S PACE C OMPLEMENTARY F ILTERS
The complementary filter can be interpreted in state 5
space form in order to easily realize the complementary filter
Orientation (deg)
with digital systems, implement multiple-input multiple- 0
output systems, and use classical control theory for filter
design [6], [23]. Figure 14 illustrates this conversion with −5
an example of the frequency domain complimentary filter
where φi and φg are the outputs from the inclinometer and −10
the gyro, Gi (s) and Gg (s) are the filters for the inclinometer
and gyro, and φ̂ is the estimate of φ. −15
0 50 100 150
Time (s)
To be a complementary filter, Gi (s) + Gg (s) must equal
1 (as shown in Eq. (2)). Therefore Gi (s) and Gg (s) could
be set as the following: Fig. 17. Estimated orientation vs. actual orientation with P = 0.84
(Optimal Gain).

C(s)
Gi (s) = , (6)
C(s) + s
ẋ = (φi − x)P + φg . (9)
s
Gg (s) = . (7)
C(s) + s A. SSCF Example - Using a Gain Controller
According to the diagram in Fig. 14 and the definitions To test the SSCF from Eq. (9), the simulator from Sec-
of Gi (s) and Gg (s), φ̂ is defined as tion II-A was used. After running this simulation multiple
times for different values of P , they are plotted with their
respective error variance to find the gain P with the lowest
C(s) φg s error. Figure 16 shows this plot and a minimum variance of
φ̂ = φi + , σ 2 = 4.15 with a gain of 0.84. Figures 17 and 18 show the
C(s) + s s C(s) + s
 1 filter output and squared error plots using the optimal gain.
φ̂ = φi C(s)s + φg − C(s)φ̂ ,
s Even though it was subjected to the same input, the
sφ̂ = (φi − φ̂)C(s) + φg . (8) performance of this filter was not as good as the frequency
domain filter designed in Section II. This could be because
Equation (8) represents the frequency domain comple- of constant bias in the gyro. Introducing an integrator in the
mentary filter in the closed-loop form as displayed in Fig. controller could improve the performance by removing this
15. With this design, classical closed-loop control theory can bias.
be used to design the complementary filter.
B. SSCF Example - Using a PI Controller
After choosing a controller, Eq. (8) can be converted to
state space. Using a gain controller, P , Eq. (9) shows the After adding an integrator to the controller C(s), Eq. (8)
state space representation of a complementary filter with φ̂ can be written as shown below where P is constant gain and
set as the state space variable x. I is integrator gain.
504
 2
Estimation Error: P=0.84, =4.1502 Variance vs. Gains
35 0.4

0.35 2.9
30

0.3
25 2.8

0.25
2.7
Error Squared

20

IGain
0.2
2.6
15
0.15
2.5
10
0.1
X: 0.44
Y: 0.05
Z: 2.27 2.4
5 0.05

2.3
0 0
0 50 100 150 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6
Time (s) PGain

Fig. 18. Squared error for P = 0.84 (Optimal Gain). Fig. 20. Variance of the error vs. controller gains (P and I) for SSCF.

Estimated Output vs. Actual Orientation: P=0.44, I=0.048

15
Actual Orientation
Estimated Output

10

Fig. 19. Simulink Model of State Space Complementary Filter. 5

Orientation (deg) 0

sφ̂ = (φi − φ̂)(P + I/s) + φg , −5

sφ̂ = (φi − φ̂)P + (φi − φ̂)I/s + φg . −10

(10)
−15
0 50 100 150
The term (φi − φ̂)I/s is compensation for the bias and Time (s)

can be set as one of the state variables x2 . x1 will remain

equal to the estimated angle φ̂. Fig. 21. Estimated orientation vs. actual orientation with P = 0.44 and
I = 0.048 (Optimal PI Gains).

x˙1 = (φi − x1 )P + x2 + φg , Estimation Error: : P=0.44, I=0.048,

2
=2.2696
20
x˙2 = (φi − x1 )I. (11)
18

16
Notice the similarities between Eq. (11) and Eq. (9).
Equation (11) can now be rearranged into its matrix repre- 14

sentation (Eq. (12)) where b is the gyro bias. The -1 found 12

Error Squared

in the ‘C’ matrix is negative in order to cancel out the gyro

10
bias.
8

6
−P 1 x1
       
x˙1 P 1 φi
= + , 4
x˙2 −I 0 x2 I 0 φg
" #  2
 
φ̂ 1 0 x1 0
= . (12) 0 50 100 150
b̂ 0 −1 x2 Time (s)

Figure 19 shows the realization of Eq. (12) in Simulink. Fig. 22. Squared error for P = 0.44 and I = 0.048 (Optimal PI Gains).
This model was used with the same input signal and sensor
models from the simulation in Section II-A. Multiple values
of P and I were simulated, the variance of the error was Figures 21 and 22 show the output of the filter with
found, and the contour plot in Fig. 20 was generated to find optimal gains. With a variance of σ 2 = 2.31, the perfor-
the optimal values for P and I that minimize the variance mance of this filter is very similar to the performance of
of the error. As shown in Fig. 20, the optimal values for P the frequency domain filter designed in Section II. This
and I in this example are 0.44 and 0.048, respectively, with similarity is expected. If the complementary filter from
a variance of σ 2 = 2.31. Eq. (12) were converted back into its frequency domain
505
 1.6
Estimated Bias vs. Actual Bias: P=0.44, I=0.048 Although Gaussian statistical models are traditional and
Actual Bias
Estimated Bias
have a wide range of applications, non-Gaussian, power-law
1.4
statics are ubiquitous in practice [27]. Although augmented
1.2 Kalman-style filters (such as the Kalman-Levy filter) that
1
allow non-Gaussian PDFs have been considered in research,
they are not well studied and have even higher mathematical
Bias (deg)

0.8 complexity than standard Kalman techniques [28].

0.6
Complimentary filters work without assuming Gaussian
0.4 noise statistics and are applicable to a wider variety of sen-
0.2
sor hardware such as commonly-available MEMS devices
with longer-tail noise tendencies [29]. Since complementary
0
filters require less computational power, they couple well
−0.2
0 50 100 150
with low-cost MEMs sensors and microcontrollers and can
Time (s) be better suited to the navigation needs of small, low-cost
UASs.
Fig. 23. Estimated Bias Plots for P = 0.44 and I = 0.048.
Our suggested guidelines include (1) understand each
sensor’s pros and cons; (2) understand the tradeoff objec-
tives; (3) whenever possible, as demonstrated in this paper,
form shown in Fig. 14, the system would be identical to
try to iteratively decide the best cut-off frequency; and (4)
the second-order filters shown in Eqs. (3) and (4), where
use as much domain knowledge as possible when deciding
P = 2ω0 and I = ω02 . Therefore, these two systems are
the design parameters linked to the complementary filter
equivalent. However, the state space system would be easier
design.
to implement using a digital computer. In addition, it is
possible to estimate the bias in the state space system since it
is added as an additional state. Figure 23 shows the estimated V. C ONCLUSION
bias vs. the actual bias.
This paper has shown how complementary simple filters
C. Other SSCF Design Options can be used to combine the output of two sensors in order
to generate an improved estimate. A frequency domain
Since the SSCF is derived using the closed-loop control complementary filter was introduced to give background on
version of the complementary filter, any other type of how complementary filters were first created. An example
controller could be used to help improve the performance of was given and simulated to show how the filter operates and
the filter. For example, a full PID controller could possibly the importance of selecting the optimal cutoff frequency.
perform better in some systems than the PI controller. Other Then the state space complementary filter was introduced
possibilities include model based controllers and fractional and derived from the definition of the frequency domain
order controllers. complementary filter. More examples were given and simu-
lated to show how this filter could be designed to estimate
SSCFs are also used for multiple-input multiple-output and eliminate gyro bias with a PI controller. This filter
systems. These include SSCFs to estimate three-dimensional preformed as well as the second-order frequency domain
orientation using Euler-angles, quaternions [16], and direct complementary filter and it was shown that they are actually
cosine matrices [20], [24]. Work has also been done on equivalent. However, the state space complementary filter
estimating position and velocity using complementary filters has advantages in digital and multiple-input multiple-output
[25]. systems. In addition, classical closed-loop control theory can
be used to design the state space filter. Overall, this paper
IV. G UIDELINES FOR C OMPLEMENTARY F ILTER U SE IN can be used to help one obtain the background necessary to
S MALL UAS understand more complex complementary filters.
Kalman or Kalman-style combining filters are the estab-
lished solution for combining sensor data into navigation- R EFERENCES
ready data. Extended Kalman filter approaches allow non-
[1] M. Bryson and S. Sukkarieh, “Vehicle Model Aided Inertial
linear models to be used and estimated with the Kalman Navigation for a UAV Using Low-cost Sensors,” in Proc. of the
techniques; however, they have proved difficult to apply to Australasian Conference on Robotics and Automation, 2006.
small UAS with low-cost, high-noise sensors [3]. Even EKF- [2] J. L. Crassidis, L. F. Markley, and Y. Cheng, “Survey of Nonlinear
based techniques are known to give unsatisfactory results Attitude Estimation Methods,” Journal of Guidance, Control, and
in these applications [26]. Complementary filters are not Dynamics, vol. 30, no. 1, pp. 12–28, 2007.
mathematically rigorous like the Kalman filter, and are [3] M. Jun, S. I. Roumeliotis, and G. S. Sukhatme, “State Estimation
therefore not well-suited for high-risk applications like space of an Autonomous Helicopter Using Kalman Filtering,” in Proc. of
borne missions. However, compared to Kalman filtering, the IEEE/RSJ International Conference on Intelligent Robots and
Systems, vol. 3, 1999, pp. 1346–1353.
complementary filter techniques are less computationally
[4] D. B. Kingston and A. W. Beard, “Real-time Attitude and Position
intensive and more readily performed on small, low-power Estimation for Small UAVs using Low-cost Sensors,” in Proc. of
computing hardware, which is ideal for small, low-cost the AIAA Unmanned Unlimited Technical Conference, Workshop
UASs. and Exhibit, 2004, pp. 2004–6488.

506
 [5] H. Rehbinder and X. Hu, “Drift-free Attitude Estimation for Measurements: A Theoretical Study,” in Proc. of the IEEE Confer-
Accelerated Rigid Bodies,” Automatica, vol. 40, no. 4, pp. 653–659, ence on Decision and Control held jointly with the Chinese Control
Apr. 2004. Conference., 2009, pp. 6407–6412.
[6] R. Mahony, T. Hamel, and J.-M. Pflimlin, “Complementary Filter [25] G. Baldwin, R. Mahony, J. Trumpf, T. Hamel, and T. Cheviron,
Design on the Special Orthogonal Group SO(3),” in Proc. of the “Complementary Filter Design on the Special Euclidean Group
IEEE Conference on Decision and Control and European Control SE(3),” in Proc. of the European Control Conference, vol. 1, no. 3,
Conference, 2005, pp. 1477–1484. 2007, pp. 3763–3770.
[7] W. Higgins, “A Comparison of Complementary and Kalman [26] J. M. Roberts, P. I. Corke, and G. Buskey, “Low-cost flight control
Filtering,” IEEE Transactions on Aerospace and Electronic Systems, system for a small autonomous helicopter,” in Proc. of the Aus-
vol. AES-11, no. 3, pp. 321–325, May 1975. tralasian Conference on Robotics and Automation, vol. 1, 2002, pp.
[8] A.-J. Baerveldt and R. Klang, “A Low-cost and Low-weight Attitude 71–76.
Estimation System for an Autonomous Helicopter,” in Proc. of the [27] M. Newman, “Power laws, pareto distributions and zipfs law,”
IEEE International Conference on Intelligent Engineering Systems, Contemporary Physics, vol. 46, no. 5, pp. 323–351, 2005.
1997, pp. 391–395. [28] N. Gordon, J. Percival, and M. Robinson, “The kalman-levy filter
[9] J. M. Roberts, P. I. Corke, and G. Buskey, “Low-cost Flight Control and heavy-tailed models for tracking maneuvering targets,” in Proc.
System for a Small Autonomous Helicopter,” in Proc. of the IEEE of the International Conference on Information Fusion, 2003, pp.
International Conference on Robotics and Automation, vol. 1, 2003, 1024–1031.
pp. 546–551. [29] N. I. Krobka, “Differential methods of identifying gyro noise struc-
[10] T. Hamel and R. Mahony, “Attitude Estimation on SO[3] Based on ture,” Gyroscopy and Navigation, vol. 2, no. 3, pp. 126–137, 2011.
Direct Inertial Measurements,” in Proc. of the IEEE International
Conference on Robotics and Automation, 2006, pp. 2170–2175.
[11] R. G. Brown and P. Y. C. Hwang, Introduction to Random Signals
and Applied Kalman Filtering, 2nd ed., ser. Wiley, TK5102.5.B696.
John Wiley & Sons, 1997, no. 4.
[12] W. H. Wirkler, “Aircraft Course Stabilizing Means,” pp. 23, Patent
2,548,278, 1951.
[13] W. Anderson and E. Fritze, “Instrument Approach System Steering
Computer,” Proc. of the Institute of Radio Engineers, vol. 41, no. 2,
pp. 219–228, Feb. 1953.
[14] F. R. Shaw and K. Srinivasan, “Bandwidth Enhancement of Position
Measurements Using Measured Acceleration,” Mechanical Systems
and Signal Processing, vol. 4, no. 1, pp. 23–38, 1990.
[15] M. Zimmermann and W. Sulzer, “High Bandwidth Orientation Mea-
surement and Control Based on Complementary Filtering,” in Proc.
of the SYROCO IFAC Symposium on Robot Control, Vienna, Austria,
Sep. 1991.
[16] E. R. Bachmann, I. Duman, U. Y. Usta, R. B. McGhee, X. P. Yun,
and M. J. Zyda, “Orientation Tracking for Humans and Robots Using
Inertial Sensors,” in Proc. of the IEEE International Symposium on
Computational Intelligence in Robotics and Automation, 1999, pp.
187–194.
[17] E. R. Bachmann, R. B. McGhee, X. Yun, and M. J. Zyda,
“Inertial and Magnetic Posture Tracking for Inserting Humans into
Networked Virtual Environments,” in Proc. of the ACM Symposium
on Virtual Reality Software and Technology (VRST). ACM Press,
2001, p. 9.
[18] E. R. Bachmann, D. McKinney, R. B. McGhee, and M. J.
Zyda, “Design and Implementation of MARG Sensors for 3-DOF
Orientation Measurement of Rigid Bodies,” in Proc. of the IEEE
International Conference on Robotics and Automation, 2003, pp.
1171–1178.
[19] S. Salcudean, “A Globally Convergent Angular Velocity Observer
for Rigid Body Motion,” IEEE Transactions on Automatic Control,
vol. 36, no. 12, pp. 1493–1497, 1991.
[20] J. M. Pflimlin, T. Hamel, and P. Souères, “Nonlinear Attitude
and Gyroscope’s Bias Estimation for a VTOL UAV,” International
Journal of Systems Science, vol. 38, no. 3, pp. 197–210, Jan. 2007.
[21] R. Mahony, T. Hamel, and J.-M. Pflimlin, “Nonlinear Complemen-
tary Filters on the Special Orthogonal Group,” IEEE Transactions
on Automatic Control, vol. 53, no. 5, pp. 1203–1218, 2008.
[22] A. R. Plummer, “Optimal Complementary Filters and Their
Application in Motion Measurement,” Institution of Mechanical
Engineers, Part I: Journal of Systems and Control Engineering, vol.
220, no. 6, pp. 489–507, 2010.
[23] P. Oliveira, I. Kaminer, and A. Pascoal, “Navigation System Design
Using Time-varying Complementary Filters,” IEEE Transactions on
Aerospace and Electronic Systems, vol. 36, no. 4, pp. 1099–1114,
2000.
[24] R. Mahony, T. Hamel, J. Trumpf, and C. Lageman, “Nonlinear
Attitude Observers on SO(3) for Complementary and Compatible

507