I.

INTRODUCTION
Attitude estimation in unmanned aerial vehicles (UAVs)
enables proper stability, control, and navigation and ensures
Cooperative Relative UAV the success of imaging and tracking missions. Aircraft
attitude consists of three degrees-of-freedom (DoFs), spec-
Attitude Estimation Using DoA ified using yaw, pitch, and roll (YPR) angles, orientation
matrices, or quaternions [1]. A variety of approaches ex-
and RF Polarization ist for estimating attitude, with the appropriate selection
depending on size, weight, cost, and accuracy constraints.
This article focuses on the problem of relative attitude
estimation, whose solution can serve as a required input
for cooperative navigation and control as well as collec-
JON W. WALLACE , Senior Member, IEEE tive localization in robot swarms [2]–[6]. Relative attitude
Lafayette College, Easton, PA, USA of ground-based robots only requires bearing information,
ATTIYA MAHMOOD, Student Member, IEEE which can be obtained with sonar, laser sensing, cameras,
Wavetronix LLC, Provo, UT, USA etc. However, obtaining the full three-DoF relative attitude
of UAVs is more challenging.
MICHAEL A. JENSEN , Fellow, IEEE
Brigham Young University, Provo, UT, USA
An obvious solution for obtaining attitude information
is for the UAVs to individually estimate their attitude with
RASHID MEHMOOD, Member, IEEE respect to a common frame of reference (such as the earth)
Wavetronix LLC, Provo, UT, USA
and, then, share that information wirelessly. Traditionally,
manned and unmanned aircraft use inertial tracking along
with the compass information to determine attitude [7]–
Relative unmanned aerial vehicle attitude is estimated using only [9]. However, disadvantages of these techniques for UAVs
on-board radio-frequency signaling. The method uses a direction-of- include poor accuracy of small low-cost inertial sensors,
arrival (DoA) vector estimate to determine two degrees-of-freedom accelerometer error from high accelerations, gyroscope
(DoFs), a polarimetric narrowband multiple-input multiple-output accumulation error, and magnetometer error due to local
(MIMO) channel estimate to specify the third DoF to within a 180◦
ambiguity, and one of the several potential methods for ambiguity
disturbances of the earth’s magnetic field. Alternatively,
resolution. Simulation results demonstrate that the method accurately it is possible to estimate attitude by optically tracking
determines aircraft attitude, with errors proportional to DoA and visual references, such as the horizon, sun, stars, or land-
MIMO channel estimate errors. This approach is useful for coopera- marks [10]–[14]. While optical methods have demonstrated
tive navigation when external navigation aids are not available, such acceptable accuracy, they are limited to conditions with
as in GPS-denied environments.
sufficient visibility.
Global positioning system (GPS) technology overcomes
most difficulties associated with other technologies, allow-
ing estimation of both location and attitude using multi-
ple receive antennas [15]–[18]. However, GPS signals are
susceptible to obstruction of the required satellites or in-
tentional jamming, and typical GPS-only methods provide
only two of the three attitude DoFs. Full three-DoF atti-
tude estimation is possible by employing widely separated
GPS receive antennas [15], but such a configuration is not
appropriate for many smaller UAVs. For these reasons,
Manuscript received January 17, 2019; revised May 11, 2019 and August fusion methods have been developed that combine GPS
29, 2019; released for publication October 30, 2019. Date of publication with existing non-GPS sensors to overcome limitations and
November 11, 2019; date of current version August 7, 2020. improve performance [19], [20].
DOI. No. 10.1109/TAES.2019.2952701 Given the drawbacks of existing approaches, particu-
larly for GPS-denied scenarios [21], an attractive alternative
Refereeing of this contribution was handled by O. Osechas.
is to use radio-frequency (RF) techniques for attitude de-
Authors’ Addresses: J. W. Wallace is with the Lafayette College, termination, possibly leveraging existing communications
Easton, PA 18042 USA, E-mail: ([email protected]); A. Mahmood and radios, and eliminating additional sensors. RF methods have
R. Mehmood are with the Wavetronix LLC, Provo, UT 84606 USA, E-mail:
([email protected]; [email protected]); M. A. Jensen
a long history of use for navigation in aviation, with systems
is with the Department of Electrical and Computer Engineering, Brigham such as very high frequency omnidirectional range and
Young University, Provo, UT 84602 USA, E-mail: ([email protected]). nondirectional beacon (NDB) stations still in active use.
(Corresponding author: Jon W. Wallace.) Although these systems are designed for localization, the
work in [22] and [23] demonstrates the use of low-frequency
0018-9251 © 2019 IEEE (LF) radio waves (available from NDBs and amplitude

IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020 2689

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 modulation (AM) radio stations) for three-DoF attitude TABLE I
estimation by exploiting the polarization of LF signals. Mathematical Notation
The work in [24] shows the use of monopulse methods
to estimate the three-DoF attitude of a satellite using RF
signals from at least two ground stations. In [25], multiple
antennas on both the UAV and a ground station enable
three-DoF attitude estimation by estimating the relative path
lengths between all antenna pairs. Finally, satellite attitude
estimation using multiple low earth orbit satellites is treated
in [26]. However, the need for dedicated ground or space-
based equipment, the susceptibility to signal obstruction
and outage, and in some cases the need for ultrawideband
(UWB) signals and station/aircraft proximity makes these
methods impractical for many applications.
Some methods have appeared that provide relative at-
titude estimates without the need for dedicated ground or
space equipment. The technique described in [27] estimates
the relative attitude of two spacecraft using at least two
transmit antennas on one vehicle and three receive antennas
on the other based on a method similar to that in [25], but
accuracy requires UWB signaling and vehicle proximity. Fig. 1. Problem scenario showing how local right-hand coordinate
systems are defined for tracking (T) and handoff (H) UAVs. The vector k̂
The methods in [28] and [29] use narrowband direction gives the direction-of-arrival (DoA) at the handoff UAV for a signal
finding for relative attitude estimation but offer only two transmitted from the tracking UAV.
DoFs, requiring use of GPS signals or a magnetometer to
obtain the third. Other related work employs radio rang- being closely tied to the accuracy of the direction finding
ing combined with optical direction finding and attitude and MIMO channel estimates.
estimation for space missions [30] or uses multiple passive Table I defines the key mathematical notation adopted
RF identification tags to estimate an object’s attitude [31]. in this article. The framework for orthogonal coordinate
In [32], a method is presented that fuses data from iner- systems and transformations is equivalent to that found
tial measurement units (IMUs), magnetometers, downward in [36], but we use alternate notation more typical of signal
looking cameras, and RF ranging to provide high-quality processing and electromagnetics literature. For those more
estimation of relative UAV pose. Similarly, the work in [33] familiar with the conventions in [36], Table I shows how to
provides a method for position, velocity, and attitude esti- translate notation.
mation for a group of miniature air vehicles by combining
IMU, range, and bearing information of nearest neighbors. II. PROBLEM GEOMETRY
Unfortunately, these existing methods do not provide a uni- In this article, we adopt the scenario for relative attitude
fied RF method that can estimate relative three-DoF attitude estimation found in [21] and depicted in Fig. 1. The tracking
using moderate bandwidth for small UAVs at arbitrary range UAV (UAVT ) currently performs some operation (such as
without requiring additional sensors. imaging of a ground target), and this UAV must be replaced
The purpose of this article is to show that by increas- by a handoff UAV (UAVH ) that takes over the operation.
ing the complexity of RF resources on a UAV (multiple Successful handoff requires that the handoff UAV learn its
antennas and multichannel radios), full relative attitude three-DoF attitude relative to that of the tracking UAV. We
of two UAVs can be estimated without the need of any could pose the problem as UAVH directly estimating the
non-RF sensors. The method uses direction finding to ob- required rotation (YPR) angles to align itself with UAVT .
tain two attitude DoFs and narrowband, multipolarization However, rotation angles depend on the order of rotation
multiple-input multiple-output (MIMO) channel estimates and, therefore, do not specify attitude unambiguously. Fur-
to resolve the third. The method has the advantages over thermore, coordinate transformations based on YPR an-
existing polarization-based RF attitude estimation [22], [23] gles can be complicated, leading to a difficult estimation
that no ground-based infrastructure is required, that anten- problem.
nas operating at UHF and above are likely already to be To avoid these difficulties, we instead take an orientation
available on UAVs, and that the antennas may be smaller matrix approach by defining the attitude of each UAV in
than those required for LF systems. While initial simulation terms of three unit vectors that form a local right-hand
results based on this algorithm appear in prior work [34], orthogonal coordinate system (or frame), as depicted in
[35], these reports omit mathematical derivation, algorithm Fig. 1. As there are several frames involved, we use the
description, and the detailed simulation results provided in general notation
this work. These simulation results demonstrate that the  A A A
algorithm provides accurate attitude estimates, with errors x̂B , ŷB , ẑB (1)

2690 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 as orthogonal unit vectors defining a Cartesian frame, where
the subscript “B” indicates the frame used as the basis for
the vector elements and the superscript “A” specifies the
new frame defined by these vectors. No subscript means
that the vector elements are in terms of the global basis
vectors (x̂, ŷ, ẑ ), which could, for example, refer to direc-
tions of increasing longitude, latitude, and altitude. Frame
A ∈ {T, H} indicates the tracking and handoff UAV body Fig. 2. KVU reference frames of the tracking and handoff UAVs are
frame, respectively, where x̂ A , ŷA , and ẑA are directions of related by a simple rotation of angle α about the common DoA vector.
the nose, left wing, and up from the fuselage, respectively, Thus, a vector in the handoff frame must be rotated by −α to represent it
expressed in the global frame. Note that the frame and in the tracking frame.
basis labels can also be identical, such as (k̂AA , v̂AA , ûAA ),
which indicates an alternate frame for A defined using the to send the resulting attitude estimate to the other
underlying Cartesian basis (x̂ A , ŷA , ẑA ). When Cartesian UAV.
unit vectors have the same subscript and superscript, they 5) All inputs to the algorithm (MIMO channel and local
are simply elementary vectors, such as ẑAA = [0 0 1]† . When DoA estimates at the two UAVs) are synchronized in
needed, roll, pitch, and yaw are defined as positive rotation time. If the MIMO channel is used for DoA estima-
angles according to the right-hand rule about the x̂ A , ŷA , and tion (see Section IV-F), the two DoA estimates and
ẑA axes, respectively. the MIMO channel are inherently synchronized. If
To simplify shifting from one frame to another, we a unified DoA/MIMO channel estimation procedure
define MAB to be the transformation matrix from A to B is not used, any relative delay in DoA and MIMO
that satisfies estimates will increase attitude estimation errors.
rB = MBA rA (2)
Assumption 3 may be difficult to meet in practice,
where rA and rB are the same vector defined in terms of the requiring patterns to be carefully measured in an anechoic
bases of frames A and B, respectively. We have chamber or in the air. Unfortunately, these patterns may be
   † corrupted by moving control surfaces, changing payload
MBA = x̂BA ŷBA ẑBA = x̂AB ŷAB ẑAB . (3)
on the outside of the UAV, and multipath effects. Although
Note that MAB = (MBA )−1 = (MBA )† since we use orthonor- beyond the scope of this first treatment, we expect that
mal basis vectors. careful transformation or processing of MIMO channel
information will lead to algorithms that require less detailed
III. PROBLEM SOLUTION knowledge and are more robust to pattern uncertainty.
Before tackling the solution, let us first state assumed For analytical convenience, we assume that no esti-
knowledge (inputs) and capabilities of the UAV systems. mation error exists in the local DoA estimates or in the
polarimetric MIMO channel estimates used in the following
1) An estimate of the MIMO channel (Hmeas ) between derivation. However, both DoA and channel estimation
the tracking and handoff UAV arrays is available. error are included in the simulations in Section IV.
Sending training symbols in one direction requires
time tmeas = NT / fs , where NT is the number of trans- A. Relating Frames With DoA Information
mit antennas and fs is the symbol rate. It should be
ensured that 1/tmeas is large compared to the rate We begin by defining a local KVU frame in each UAV
change of relative attitude. local frame by the unit vector triad (k̂AA , v̂AA , ûAA ) where
2) At least one UAV has estimates of k̂TT and k̂HH , which  A
(ẑA × k̂AA )/ẑAA × k̂AA , for k̂AA = ẑAA
represent the direction-of-arrival (DoA) vector in ûAA = (4)
x̂AA , otherwise
each UAV local frame. This local DoA information
does not require GPS, since UAVs can use arrays for v̂AA = ûA
A
× k̂AA (5)
direction finding [37], [38].
3) UAVs have knowledge of the complex (amplitude and A ∈ {T, H}. The KVU unit vectors are analogous to
and phase) polarimetric far-field radiation patterns spherical unit vectors. The KVU frames for UAVT and
of their antenna arrays over the complete radiation UAVH are denoted as t and h, respectively. The vectors k̂ T
sphere. Although this detailed knowledge would be and k̂ H denote the DoA vector k̂ as seen by the tracking and
sufficient to estimate and/or correct antenna phase handoff UAVs in the global frame (see Fig. 1). Since k̂ T
center differences, such operations are not required should equal k̂ H , we have the situation depicted in Fig. 2,
in our solution. where the two KVU frames are related by the simple rotation
4) A data channel exists between the UAVs to allow ⎡ ⎤
1 0 0
required information (the MIMO channel estimate, ⎢ ⎥
local DoA, antenna patterns) to be sent to the pro- Mht (α ) = ⎣ 0 cos α − sin α ⎦ (6)
cessing UAV node. This channel would also be used 0 sin α cos α

WALLACE ET AL.: COOPERATIVE RELATIVE UAV ATTITUDE ESTIMATION USING DOA AND RF POLARIZATION 2691

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 where α is an unknown that cannot be deduced from DoA With patterns in a common frame, we can compute the mnth
information alone. Conceptually, the v̂ -û axes in each frame entry of the channel transfer matrix H as
are related by a single rotation angle α. Once α is known,    
the two UAV reference frames are related by Hmn (α ) = eHTn · eHHm exp jψmn (16)

MHT (α ) = MHh Mht (α )MtT . (7) where ψmn is the RF phase shift from the nth tracking
antenna to the mth handoff antenna, computed as
The matrices MHh and MtT can be readily computed using  
(4) and (5) for A ∈ {T, H} and forming ψmn = k0 rTTn · k̂TT − rHHm · k̂HH (17)
 † k0 = 2π f /c is the wavenumber (for RF frequency f and
MtT = k̂TT v̂TT ûTT (8)
 H H H speed of light c), and rTTn and rHHm are the position (assumed
MH = k̂H v̂H ûH
h
(9) phase center) of the nth tracking and mth handoff antennas
which only requires local knowledge of k̂. in their local UAV Cartesian frames, respectively.
We find α by searching for the value that gives the best
B. Finding α Using the Polarimetric MIMO Channel fit between the measured MIMO channel Hmeas and the
modeled value given in (16). We first stack the channels
We now formulate a method for determining α by
columnwise into column vectors, with the operation being
comparing measurements of the narrowband polarimetric
denoted as
MIMO channel between the two UAVs to a model of
the channel. UAVT and UAVH have arrays with NT and hmeas = Vec{Hmeas } (18)
NH antennas, respectively. It is most natural to define the h(α ) = Vec{H(α )}. (19)
field radiation pattern of each antenna element using the
local frame aligned with that antenna. We, therefore, define Since it is difficult to measure or estimate absolute ampli-
Cartesian frames for the nth tracking and mth handoff tudes and phases of the channel matrix entries, we normalize
antennas as (x̂TTn , ŷTTn , ẑTTn ) and (x̂HHm , ŷHHm , ẑHHm ), referred to the channel vectors according to
as the Tn and Hm frames, respectively. Patterns are naturally
hmeas = hmeas /hmeas, (20)
defined in spherical KVU frames given by (k̂TTnn , v̂TTnn , ûTTnn )
and (k̂HHmm , v̂HHmm , ûHHm
m
), referred to as the t n and hm frames, h(α ) = h(α )/h (α ) (21)
respectively. where  = arg max |hmeas, |. Note that noise amplification
It is assumed that each antenna is simulated or measured due to this normalization should be small since the chan-
in its local frame, such that the electric field patterns etTnn (k̂TTnn ) nel element with the maximum amplitude is always used.
and ehHmm (k̂HHmm ) are known. Note that for far-field patterns, Finally, α is given by the solution of
the vector field expressed in the spherical basis is [0 eθ eφ ]† ,
since field polarized in the radiation direction is zero. αopt = arg min hmeas − h(α )2 . (22)
α
Modeling the channel between the two UAVs requires
the pattern in the direction of k̂ defined in each antenna Since we are only seeking a single real number over the
frame, which is given by interval [−π, π ) in (22), a simple one-dimensional (1-D)
parameter sweep can be used. The relation between the T
k̂TTnn = MTTn k̂TT (10) and H frames can be found using (7), or
k̂HHmm = −MHHm k̂HH (11)  †  
MHT (αopt ) = x̂TH ŷTH ẑTH = x̂HT ŷHT ẑHT . (23)
where the minus sign ensures that k̂HHmm points radially away
Because absolute phase is unavailable, the phase normaliza-
from UAVH for the definition of the handoff UAV antenna
tion in (20) and (21) creates a 180◦ ambiguity in α, meaning
patterns. Note that the direction k̂ in a KVU reference frame
MHT (αopt + π ) is also a candidate solution.
can be converted to spherical angles using
 
φ = tan−1 ky /kx (12) C. Resolving the 180◦ Ambiguity
 
θ = tan−1 kx2 + ky2 /kz (13) Several methods for resolving the 180◦ ambiguity in α
are described as follows.
where (kx , ky , kz ) are the Cartesian components of k̂, and φ 1) Multiple Aspects: Appendix A proves that applying
and θ are the usual spherical azimuth and elevation angles. the method for two different values of k̂ but for the same
With UAVT as the transmitter, the vector (polarized) relative attitudes gives the same attitude estimates for the
field generated by the nth antenna in the reference frame correct value of α but different attitude estimates for the
for UAVH is given by solution α + π. Removing the incorrect solution requires
maneuvering to change the relative positions but not the
eHTn = MHT (α )MTTn MTtnn etTnn . (14) relative attitudes of the UAVs. Linear motion is an example
With UAVH as the receiver, the polarized reception pattern of such a maneuver. This technique is appropriate for UAVs
of the mth antenna is that are close enough for the global k̂ vector to change
significantly (perhaps a few degrees) over a short time,
eHHm = MHHm MHhmm ehHmm . (15) which is not the case for widely separated UAVs.

2692 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 2) Tracking: Given known initial relative attitudes, the 5) Generate samples of α on the interval [0, π ], and
method is applied often enough to allow comparison of for each candidate value of α perform the following
subsequent attitude estimates. Using the correct value of steps.
α leads to correct attitude estimates that gradually change a) Compute Mht (α ) using (6).
over time. Using the 180◦ ambiguity leads to a large jump b) Compute MHT (α ) using (7).
in the attitude (typically 180◦ ). In this way, as the relative c) Compute the entries of the modeled MIMO
pose is successively tracked, the α giving an attitude that channel by performing the following steps for
is closest (such as in a least squared sense) to the previous all pairs of m and n.
step is retained as the correct solution. With the exception of i) Construct MTTn = (MTTn )† and
obtaining the initial relative attitude estimate, it is expected MHm = (MH ) based on known
H Hm †

that this technique could be used in virtually any scenario. orientations of the antennas relative
3) Attitude Constraints: For most UAVs, the range of to each local UAV frame.
attitudes that are consistent with controlled flight is limited. ii) Find k̂TTnn and k̂HHmm in the local antenna
Often, the YPR associated with the solution α + π produces KVU frames of UAVT and UAVH ,
nonsensical attitudes, such as an overly steep roll or pitch respectively, using (10) and (11).
angle or even a UAV flying upside down, allowing identi- iii) Look up the polarimetric patterns of
fication and rejection of the incorrect value of α. This kind the nth UAVT and mth UAVH anten-
of sanity checking should be used in conjunction with other nas by converting the local k̂ direc-
methods, since not all ambiguities can be resolved this way. tions from the last step into spher-
4) Path-Length Information: Methods such as those ical angles using (12) and (13). To
presented in [25] and [27] use relative path distances be- allow a unique one-to-one mapping
tween pairs of transmit and receive antennas to obtain over the full [−π, π ) radian interval,
unambiguous attitude estimates. Unfortunately, those meth- tan−1 (a/b) should be computed us-
ods require very high bandwidth to obtain high accuracy, ing a two-argument tangent function
especially at long range. However, if path-length methods (hereafter designated as atan2(a, b)).
are only used for resolving the 180◦ ambiguity, much less iv) Use (4) and (5) to form transformation
accuracy (and, therefore, bandwidth) is required. The basic matrices from each antenna’s local
strategy in our case would be to estimate the path distances spherical frame to its Cartesian frame
between a set of transmit (tracking) and receive (handoff)  
MTtnn = k̂TTnn ûTTnn v̂TTnn (24)
antennas. Given the two candidate estimates of α, two differ-  Hm Hm Hm 
ent attitudes are obtained. For each attitude, the relative path MHm = k̂Hm ûHm v̂Hm . (25)
hm

lengths can be computed and compared with the measured v) Use (14) and (15) to transform the lo-
values, allowing the correct α to be identified. Although the cal polarimetric antenna patterns into
ability of these methods to obtain an unambiguous estimate a common frame of reference (the
in a single snapshot is attractive, it is unclear how much RF UAVH local frame).
bandwidth is needed. The development of such methods is vi) Compute the channel matrix element
left for future investigation. with (16) and (17).
5) Sensor Fusion: It is likely that most UAVs will not d) Stack and normalize the estimated channel
have to rely totally on RF resources for control and navi- matrix obtained from the previous loop ac-
gation, possibly incorporating an IMU, magnetic compass, cording to (21). Use the same normalization
cameras, etc. Although outside the scope of this work, we index as that used in the measured channel
expect that information from these sensors on the two UAVs previously.
could be combined to eliminate the 180◦ ambiguity, even e) Compare the modeled and measured MIMO
when such sensors do not provide enough information for channels according to (22).
estimation of the three-DoF relative attitude. 6) Given the solution to (22) from the previous loop,
referred to as αopt , compute MHT using αopt and
D. Algorithm
αopt + π, which are two candidate solutions giving
An algorithm applying the outlined method can be sum- the relative orientation vectors of the UAVs, as de-
marized as follows. scribed in (23).
7) Use a method to resolve the 180◦ ambiguity (see
1) Estimate the UAVT DoA vector with a direction- Section III-C) and select the correct estimate of MHT .
finding technique [38] locally at UAVT and UAVH
to obtain k̂TT and k̂HH , respectively. IV. ILLUSTRATIVE NUMERICAL EXAMPLES
2) Use (4) and (5) to obtain ûTT , v̂TT , ûHH , and v̂HH , which
give the local KVU frames for the two UAVs. A. UAV Model and Antenna Configuration
3) Form the transformation matrices MtT and MHh using A detailed UAV model was created using Autodesk
(8) and (9). Fusion 360 as depicted in Fig. 3. The UAV has a wingspan
4) Normalize the measured MIMO channel using (20). of 1.5 m and nose-to-tail length of 0.8 m. Arguably, a high

WALLACE ET AL.: COOPERATIVE RELATIVE UAV ATTITUDE ESTIMATION USING DOA AND RF POLARIZATION 2693

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 3. UAV model simulated with FDTD. Numbered cylinders are six Fig. 4. Normalized antenna gain patterns (in dB) from FDTD
monopole antennas used for DoA and relative attitude estimation. simulations in the yz plane (left) and xy plane (right). Circled and boxed
numbers indicate antenna indices. The small centered airplanes indicate
the proper orientation of each plot relative to the UAV as well as the
indexing of the antennas. Although 2-D gain pattern cuts are plotted here,
RF center frequency in the microwave or millimeter wave the algorithm uses complex polarimetric patterns over the full radiation
regime is ideal to reduce antenna size and weight. However, sphere.
such high frequencies hamper the full-wave electromag-
netic simulation of the complete UAV performed in this
Fig. 4 plots normalized antenna gain patterns for all
article. Although we use a somewhat low center frequency
six antennas in the yz and xy planes. Although the antenna
of 200 MHz in this example, resulting in large antennas,
patterns do resemble monopoles with lower radiation along
practical implementations may use higher frequencies to
the monopole axis, the UAV body distorts the patterns.
minimize antenna size.
Although optimal antenna array design is beyond our B. DoA and Channel Estimation
present scope, we expect that accurate DoA estimation over
the whole radiation sphere will require an antenna array with This section describes how noisy DoA and MIMO chan-
amplitude and/or phase variation with respect to DoA and nel estimates are obtained, which are used as inputs to the
no “dead spots” (directions where no antenna radiates effec- simulated attitude estimation algorithm.
tively). Next, three-DoF relative attitude estimation requires 1) Gaussian DoA Estimation Error: First, we focus
detecting a rotation around the DoA vector, meaning that on attitude estimation uninfluenced by algorithmic choices
the MIMO channel must change due to these rotations. If for the direction-finding method. The noisy estimate of the
usual unipolarized antennas are used, at least two antennas DoA vector is obtained by randomly rotating k̂ using
 
with distinct polarization need to be visible for each possible k̂est = k̂ x̂ K ŷK ẑK (26)
DoA.   
MK
Following the abovementioned observations, six
quarter-wave monopole antennas were placed on the UAV where the noise rotation matrix MK is constructed by gen-
as shown by the numbered cylinders in Fig. 3. Although erating zero-mean Gaussian random pitch, roll, and yaw
each antenna only senses (or excites) a single polarization, angles, each with variance σK2 , and using the expressions
three orthogonally oriented antennas are visible for each in Appendix B to convert these vectors to an orientation
DoA, allowing the full E-field polarization vector to be esti- matrix.
mated for any DoA. Finite-difference time-domain (FDTD) 2) Table Lookup DoA Estimation: Existing beamform-
simulations [39] were performed, where it was assumed ing and subspace methods [37], [38] may be used for DoA
that the UAV geometry is a perfect electrical conductor. estimation, but these methods often require a special array
Simulations were performed using a custom FDTD code structure or do not automatically handle multiple polar-
that has been employed successfully in several previous izations. To allow the same six-element antenna array to
modeling problems [40]. The FDTD simulation domain be used for both MIMO channel and DoA estimation, a
size was 3.0 × 3.6 × 2.4 m3 in the x, y, and z directions, straightforward DoA estimation procedure based on a table
respectively, and the FDTD unit cell size was 5 cm. Each lookup was developed and is described below. Drawbacks
antenna pattern was obtained by running a separate sim- of this simple method include storage requirements for
ulation with a gap voltage source placed across one of the database and a large number of required channel-table
the antennas from the monopole base to the aircraft body comparisons. Although the simple method meets our goal
and open-circuiting the other five antennas. Polarimetric of illustrating correct performance of the attitude estimation
far-field patterns were computed by integrating near fields algorithm, we hope that future work will adapt subspace-
on a box enclosing the complete UAV and using a standard based methods to this problem, likely improving algorith-
near-to-far field transformation. Far fields were stored for mic efficiency and robustness.
the complete radiation sphere for each antenna with an Given the simulated antenna patterns, we can com-
angular resolution of 5◦ in both elevation and azimuth. This pute hexp (θ, φ, a, ν ), which represents the expected antenna
coarse sampling was chosen to show that attitude estimation array complex terminal voltages for local DoA direction
is reliable even with nonideal pattern knowledge. (θ, φ ) and incident polarization having axial ratio |a| and tilt

2694 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 angle ν. Polarization handedness is specified by the sign of a
where a < 0 (a > 0) indicates left (right) hand polarization.
A finite size table was computed by sampling
θ = 0, 5◦ , 10◦ , . . . , 180◦ (27)
φ = 0, 5◦ , 10◦ , . . . , 355◦ (28)
a = −1, −0.9, . . . , 0.9, 1 (29)
ν = 0, 5◦ , 10◦ , . . . , 175◦ (30)
which consists of 37 × 73 × 21 × 36 = 2041956 entries.
Storage of the table requires 93 MB of memory for single
precision floating point numbers assuming a six-element
array.
Once the table is computed, the local DoA vector at
each UAV is estimated as follows. Assuming the tracking Fig. 5. RMS error of orientation vectors computed using 1000 random
attitude realizations as a function of DoA RMS error, where each curve is
and handoff UAVs to be the transmitter and receiver, re- for a fixed value of SNRdB . The three orientation vectors have identical
spectively, the handoff UAV finds the column of the channel statistics, which were averaged.
H having the highest energy according to hmaxH
= H(:, n0 ),
where
 rather than practical flight maneuvers, we cannot apply the
n0 = arg max |Hmn |2 . (31) methods for ambiguity resolution of α and, therefore, we
n
m simply choose the correct value of α based on the exact
In practice, absolute amplitude and phase are difficult to relative attitudes.
recover, and the signal vector and table are normalized as Using uniformly distributed random attitude angles can
result in some extreme attitudes, and for some of these
max = hmax /hmax,
hH H H
(32) situations, slight errors in actual UAV orientation can lead
hexp = hexp /hexp, (33) to relatively large errors in YPR angles. Therefore, we di-
rectly quantify the angular error of the estimated orientation
where  = arg max |hmaxH
, |. Finally, the table is searched vectors x̂HT , ŷHT , and ẑHT using the expression
for the entry giving the least-square fit, or
 H H H H  
r̂ = cos−1 (r̂est · r̂exact ) (36)
θ , φ , a , ν = arg min hHmax − hexp (θ, φ, a, ν )
(θ,φ,a,ν ) where r̂ ∈ {x̂HT , ŷHT , ẑHT }, and (· )est and (· )exact refer to esti-
(34) mated and exact quantities, respectively.
where θ H and φ H give the estimated elevation and azimuth Fig. 5 shows the RMS error in the orientation vectors
angles of the local DoA vector at the handoff UAV. In an of the estimated UAV attitude as a function of the DoA
analogous way, the process is repeated on the row of H estimation error. For infinite SNR, attitude estimation error
having the highest energy to obtain the local DoA vector at is approximately double that of the DoA estimation error,
the tracking UAV, or θ T and φ T . resulting from the combination of independent DoA error
3) Channel Estimation Error: The exact MIMO chan- at the tracking and handoff UAVs. For practical SNR, an
nel estimates are corrupted using additive zero-mean com- error floor in the attitude estimate is experienced at low
plex Gaussian noise with variance DoA error.
σC2 = 10−SNRdB /10 max |Hi j |2 (35)
ij D. Single Circling UAV
and SNRdB is the specified signal-to-noise ratio (SNR) in In this example, UAVT hovers in a fixed position at the
dB. Referencing the noise to the maximum channel gain origin while UAVH travels in a banked turn around UAVT at
removes the impact of antennas obstructed by the UAV that a constant radius. The free-space wavelength is λ = 1.5 m,
reduce the average gain and, therefore, essentially keeps the the circling radius is 150 m, the speed is 100 km/h, the
noise constant independent of relative UAV attitudes. bank angle is 30◦ , and both UAVs are at the same altitude.
Gaussian DoA error (see Section IV-B1) of 3◦ is used, and
C. Performance versus Estimation Error SNR is 20 dB. In this case, we resolve the ambiguity in α
To quantify the performance of the method with respect by rejecting relative pitch greater than 45◦ and relative roll
to error in the DoA and channel estimates, we implement greater than 90◦ . The results in Fig. 6 show that the attitude
the method for 1000 random orientations of the UAVs for estimator is able to accurately determine the YPR angles,
different levels of estimation error. Simple Gaussian DoA with RMS errors in these angles of (4.4◦ , 3.1◦ , 4.3◦ ).
estimation error is used, as described in Section IV-B1. The
E. Two Circling UAVs
sphere of all possible attitudes is covered in these simula-
tions by having yaw and roll uniform on [0◦ , 360◦ ] and pitch Next, we consider the case where both UAVT and UAVH
uniform on [−90◦ , 90◦ ]. Given that we use random attitudes fly in 30◦ banked turns (in opposite turn directions) on a

WALLACE ET AL.: COOPERATIVE RELATIVE UAV ATTITUDE ESTIMATION USING DOA AND RF POLARIZATION 2695

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 Fig. 8. CDFs of error for Monte Carlo simulations of attitude estimation
using the table lookup method for local DoA estimation. Solid and dashed
lines show error in local DoA estimates and resulting attitude estimates,
respectively. Labels on the curves indicate the simulated receive SNR.

Fig. 6. Estimates of relative YPR versus time for one UAV flying in a
circle of radius 150 m using a 30◦ bank around another fixed
(hovering) UAV. similar to those in Section IV-C. Local DoA estimates were
obtained using the table lookup method in Section IV-B2.
Fig. 8 shows the cumulative distribution functions (CDFs)
of the error for the local DoA estimates (solid curves),
defined as

 H 

k̂ = cos−1 k̂est · k̂exact
H
. (37)

Although DoA estimation performance is poor for 10 dB

SNR, acceptable performance is obtained for 15 dB or
higher, where 10◦ of error or better is obtained with prob-
ability 0.75, 0.9, and 0.97, for 15, 20, and 30 dB SNR,
respectively. Fig. 8 also shows the error of the resulting
attitude estimates (dashed curves). For low to moderate
SNR, relatively high error in the DoA estimates leads to
even higher error in the attitude estimation. However, for
high SNR, error in the attitude estimate becomes similar to
that of the DoA estimates, approaching a bound of about
5◦ , which is the resolution of the pattern lookup table. In
Fig. 7. Estimates of YPR versus time for two circling UAVs flying in
30◦ banked turns on circles of radius 150 m. The UAVs turn in opposite
practice, 5◦ of error in relative attitude may not be sufficient
directions to provide constantly changing relative YPR. for some applications, suggesting that interpolation or more
detailed patterns should be used.
A troubling aspect of Fig. 8 is the heavy-tail behavior
circle of radius 150 m at 100 km/h with an altitude difference that could lead to poor performance with a tracking filter.
of 45 m. Due to the larger relative attitudes, we use multiple The heavy tails in the error distributions are due to infre-
aspects (see Section III-C1) to resolve the 180◦ ambiguity in quent cases where the DoA vector at one of the UAVs has
α at the beginning of the maneuver and tracking thereafter. large error, leading to very high error in the resulting attitude
RMS DoA error and SNR are the same as the last example. estimate. Fortunately, these cases can usually be detected
Fig. 7 demonstrates that the method is able to accurately by checking the fidelity of the fit between the estimated
determine the YPR angles with respective RMS errors of and modeled MIMO channel, or
H = hmeas − h(αopt )2 .
(4.4◦ , 4.4◦ , and 4.1◦ ). If
H is much larger than typical values, this usually indi-
cates a poor attitude estimate. Fig. 9 explores an improved
F. Practical DoA Estimation
algorithm that discards attitude estimates having
H greater
Finally, we explore the performance of a method that than the median. Statistics of
H can be obtained through
uses a practical DoA estimation method, illustrating that simulation or collected during system operation. Fig. 9 plots
a single RF array can be used for both direction finding the CCDF of attitude estimation error for three SNR levels
and attitude estimation. 1000 Monte Carlo simulations were for the original and improved algorithm, indicating that far
performed with uniform random UAV attitude and DoA, more robust estimates (shorter tails) are obtained.

2696 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 This means that
rH() = MHh() Mht() MtT() rT (39)
rH()  = MHh() LMht() MtT() rT (40)

where  ∈ 1, 2 for the two different cases of k̂ and the de-

pendence on α is dropped in the notation since all quantities
now assume the correct value of α. In the absence of noise
and estimation error, by construction our algorithm always
gives the correct estimate of rH regardless of k̂ when α is
correct, meaning rH(1) = rH(2) , which implies

Fig. 9. Complementary cumulative distribution functions (CCDFs)

MHh(1) Mht(1) MtT(1) rT = MHh(2) Mht(2) MtT(2) rT . (41)
from Monte Carlo simulations of the original table-lookup algorithm (see
In contrast, when α + π is selected, we have
Fig. 8) indicated by “All Data,” and a modified algorithm where 50% of
the estimates are discarded (“Pruned 50%”) according to
H . Significant rH(1)  = MHh(1) LMht(1) MtT(1) rT (42)
reduction of the error tails is obtained.
= MHh(1) LMhH(1) MHh(2) Mht(2) MtT(2) rT (43)
V. CONCLUSION
where the second equality comes from using (41) with
This article presents a method for finding the relative (MHh(1) )−1 = MhH(1) . Next, from (40) and (43), we compute
attitude of two UAVs using only radio direction finding the difference of the two estimates as
and MIMO channel estimation techniques. The method is
useful in situations where GPS access and/or precise attitude rH(2)  − rH(1)  =
information from gyros/accelerometers is not available. The 
MHh(2) L − MHh(1) LMhH(1) MHh(2) Mht(2) MtT(2) rT . (44)
problem solution uses the DoA vector to estimate two
  
degrees of freedom, reducing the unknown relative attitude Q
to a single angle α that is then estimated using observation
of the polarimetric MIMO channel. Since the narrowband, Since the basis transformation matrices are always full rank,
polarimetric, far-field MIMO channel only gives enough the only way this difference can be zero for all vectors rT is
information to resolve α to within a 180◦ ambiguity, several for Q to be the zero matrix. This requires
methods based on practical considerations or controlled MHh(2) LMhH(2) = MHh(1) LMhH(1) . (45)
maneuvers are provided for ambiguity resolution.
Simulations based on the proposed algorithm for real- Using (8), (9), and L from (38), we can express (45) as
istic system parameters and flight profiles show that the †  † † 
method is able to estimate all three DoFs of UAV relative
H(1)
k̂H H(1)
k̂H − v̂H H(1) H(1)
v̂H H(1)
+ ûH H(1)
ûH
attitude. Monte Carlo simulations based on random UAV †  † † 
attitudes demonstrate that error in the orientation vectors is H(2)
= k̂H H(2)
k̂H − v̂H H(2) H(2)
v̂H + ûHH(2) H(2)
ûH .
proportional to both the error in the DoA vector estimate as
well as the SNR of the channel estimate. (46)
APPENDIX We know from properties of basis transformations that
A. Proof of Shifting k̂ Method MHh() MhH() = I, where I is the identity matrix, or
† † †
The purpose of this appendix is to prove that the 180◦ k̂HH() k̂HH() + v̂HH() v̂HH() + ûHH() ûHH() = I (47)
ambiguity in α can always be resolved by keeping the
aircraft attitudes constant while changing k̂ by a small which simplifies (46) to
amount. † †
Consider an arbitrary vector rT in the tracking frame. k̂HH(1) k̂HH(1) = k̂HH(2) k̂HH(2) . (48)
The handoff UAV estimates the same vector in the handoff
frame for two DoA vectors k̂ (1) and k̂ (2) with k̂ (1) = k̂ (2) , The only way for these outer products to be equal is for
resulting in estimates rH(1) and rH(2) when the correct α is k̂HH(1) = ±k̂HH(2) , but this either contradicts our initial as-
used and rH(1)  and rH(2)  when α + π is used. Recall that the sumption or requires a large maneuver that exchanges the
correct α aligns ûT and v̂ T with ûH and v̂ H in any common relative location of the two UAVs. Precluding the latter case,
reference frame (see Fig. 2). But when α + π is selected, we must have Q = 0.
we have The meaning of this result is that if the incorrect value of
⎡ ⎤ α is used, there exist some vectors rT that will be estimated
1 0 0 differently as k̂ changes, even though the relative UAV
Mht (α + π ) = Mht (α ) ⎣ 0 −1 0 ⎦ . (38) attitudes remain constant. However, when the correct value
0 0 −1
   of α is used, the same estimate will be obtained independent
L of k̂. A convenient set of test vectors for rT is (x̂TT , ŷTT , ẑTT ),

WALLACE ET AL.: COOPERATIVE RELATIVE UAV ATTITUDE ESTIMATION USING DOA AND RF POLARIZATION 2697

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 since the algorithm in this article already estimates those [4] V. Indelman, P. Gurfil, E. Rivlin, and H. Rotstein
in the handoff frame as (x̂HT , ŷHT , ẑHT ) or MHT . Note that it is Graph-based distributed cooperative navigation
possible for (44) to be zero for some test vectors, but since In Proc. IEEE Intl. Conf. Robot. Autom., Shanghai, China,
May 9–13, 2011, pp. 4786–4791.
Mht(2) MtT(2) in (44) is full rank, the error cannot be zero for [5] R. Sharma, S. Quebe, R. W. Beard, and C. N. Taylor
all three orthogonal vectors (x̂TT , ŷTT , ẑTT ) when the incorrect Bearing-only cooperative localization.
α is used. J. Intell. Robot. Syst., vol. 72, pp. 429–440, Dec. 2013.
[6] S. Chung, A. A. Paranjape, P. Dames, S. Shen, and V. Kumar
B. Relation of Orientation Frames to Euler Angles A survey on aerial swarm robotics
IEEE Trans. Robot., vol. 34, no. 4, pp. 837–855, Aug. 2018.
This article employs frames rather than Euler angles [7] P. Groves
of rotation, since this approach simplifies the analysis and Principles of GNSS Inertial and Multisensor Integrated
does not suffer from ambiguities due to order of rotation. Navigation Systems. Norwood, MA, USA: Artech House,
2007.
However, it may be desirable to convert from a frame to
[8] O. Woodman
Euler angles and vice versa, and this appendix provides such An introduction to inertial navigation
expressions. For this purpose, we assume the YPR order of University of Cambridge, Cambridge, U.K., Tech. Rep.
rotations, which is usually standard in aviation. Yaw (Y), UCAMCL-TR-696, 2007.
pitch (P), and roll (R) angles are defined as rotation about [9] D. Biezad
Integrated Navigation and Guidance Systems (ser. AIAA Edu-
the ẑT , ŷT , and x̂ T axes, respectively (see Fig. 1), with the
cation Series). Reston, VA, USA: Amer. Inst. Aeronaut. Astro-
positive direction given by the right-hand rule. naut., 1999.
1) YPR to Orientation: Yaw (Y) and pitch (P) define [10] S. M. Ettinger, M. C. Nechyba, P. G. Ifju, and M. Waszak
the direction of the x̂ T vector, analogous to how φ and θ Vision-guided flight stability and control for micro air vehicles
define the direction of the r̂ vector in spherical coordinates, In Proc. IEEE/RSJ Intl. Conf. Intell. Robots Syst.,
Lausanne, Switzerland, Sep. 30–Oct. 4, 2002, vol. 3,
or
pp. 2134–2140.
x̂ T = [cos(P ) cos(Y ) cos(P ) sin(Y ) − sin(P )]† . (49) [11] D. Dusha, L. Mejias, and R. Walker
Fixed-wing attitude estimation using tempoaral tracking of the
The other vectors for the case of zero roll can be found as horizon and optical flow
    J. Field Robot., vol. 28, pp. 355–372, Mar. 2011.
ŷT = ẑ × x̂ T /ẑ × x̂ T  (50) [12] D. Dusha
  Integrated vision-based attitude estimation
 
ẑT = x̂ T × ŷT /x̂ T × ŷT  (51) Ph.D. dissertation, Queensland Univ. Technol., Brisbane, QLD,
USA, 2012.
where ẑ = [0 0 1]† . Finally, roll (R) can be applied to obtain [13] A. E. R. Shabayek, C. Demonceaux, O. Morel, and D. Fofi
Vision based UAV attitude estimation: Progress and insights
ŷT = ŷT cos(R ) + ẑT sin(R ) (52) J. Intell. Robot. Syst., vol. 65, pp. 295–308, Jan. 2012.
[14] M. H. Tehrani, M. A. Garratt, and S. G. Anavatti
ẑT = −ŷT sin(R ) + ẑT cos(R ). (53) Low-altitude horizon-based aircraft attitude estimation using
UV-filtered panoramic images and optic flow
2) Orientation to YPR: Again, direction of the nose IEEE Trans. Aerosp. Electron. Syst., vol. 52, no. 5, pp. 2362–
(x̂ T ) gives pitch and yaw like in spherical coordinates, or 2375, Oct. 2016.
  [15] G. Lu, M. E. Cannon, G. Lachapelle, and P. Kielland
Y = atan2 x̂2T , x̂1T (54) Attitude determination using dedicated and nondedicated mul-
  
 2  2 tiantenna GPS sensors
P = −atan2 x̂3T , x̂1T + x̂2T (55) IEEE Trans. Aerosp. Electron. Syst., vol. 30, no. 4, pp. 1053–
1058, Oct. 1994.
[16] H. M. Peng, E. R. Chang, and L. S. Wang
where x̂iT denotes the ith element of the unit vector. The zero Rotation method for direction finding via GPS carrier phases
roll vectors ŷT and ẑT are found according to (50) and (51), IEEE Trans. Aerosp. Electron. Syst., vol. 36, no. 1, pp. 72–84,
and roll can then be computed as Jan. 2000.
 [17] L. Lau, P. Cross, and M. Steen
R = atan2 ŷT · ẑT , ŷT · ŷT . (56) Flight tests of error-bounded heading and pitch determination
with two GPS receivers
IEEE Trans. Aerosp. Electron. Syst., vol. 48, no. 1, pp. 388–404,
REFERENCES
Jan. 2012.
[1] W. F. Phillips, C. E. Hailey, and G. A. Gebert [18] P. Henkel and M. Iafrancesco
Review of attitude representations used for aircraft kinematics Tightly coupled position and attitude determination with two
J. Aircraft, vol. 38, pp. 718–737, Jul./Aug. 2001. low-cost GNSS receivers
[2] A. C. Sanderson In Proc. 11th Intl. Symp. Wireless Commun. Syst., Barcelona,
A distributed algorithm for cooperative navigation among mul- Spain, Aug. 26–29, 2014, pp. 895–900.
tiple mobile robots [19] J. N. Gross, Y. Gu, M. B. Rhudy, S. Gururajan, and
Adv. Robot., vol. 12, pp. 335–349, 1997. M. R. Napolitano
[3] S. I. Roumeliotis and G. A. Bekey Flight-test evaluation of sensor fusion algorithms for attitude
Distributed multirobot localization estimation
IEEE Trans. Robot. Autom., vol. 18, no. 5, pp. 781–795, Oct. IEEE Trans. Aerosp. Electron. Syst., vol. 48, no. 3, pp. 2128–
2002. 2139, Jul. 2012.

2698 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 [20] Z. Wu, M. Yao, H. Ma, W. Jia, and F. Tian [30] X. Wang, X. Shao, D. Gong, and D. Duan
Low-cost antenna attitude estimation by fusing inertial sens- Radio/VISNAV integrated navigation system for autonomous
ing and two-antenna GPS for vehicle-mounted satcom-on-the- spacecraft rendezvous using two-order filter
move In Proc. 8th World Congr. Intell. Control Autom., Jinan, China,
IEEE Trans. Veh. Technol., vol. 62, no. 3, pp. 1084–1096, Jul. 7–9, 2010, pp. 1735–1740.
Mar. 2013. [31] G. A. Narciandi, J. Laviada, and F. Las-Heras
[21] H. Bai and R. W. Beard Object attitude estimation using passive RFID tag arrays
Relative heading estimation and its application in target handoff In Proc. URSI Intl. Symp. Electromagn. Theory, Espoo, Finland,
in GPS-denied environments Aug. 14–16, 2016, pp. 572–574.
IEEE Trans. Control Syst. Technol., vol. 27, no. 1, pp. 74–85, [32] J. Hardy et al.
Jan. 2019. Unmanned aerial vehicle relative navigation in GPS denied
[22] S. T. G. Maguire and P. A. Robertson environments
Low frequency radio polarization sensor with applications in In Proc. IEEE/ION Position, Location Navigation Symp.,
attitude estimation Savannah, GA, USA, Apr. 11–14, 2016, pp. 344–352.
IEEE Sensors J., vol. 15, no. 12, pp. 7304–7311, Dec. 2015. [33] R. Sharma and C. Taylor
[23] S. T. G. Maguire and P. A. Robertson Cooperative navigation of MAVs in GPS denied areas
UAV attitude estimation using low-frequency radio polarization In Proc. Intl. Conf. Multisensor Fusion Integration Intell. Syst.,
measurements Seoul, South Korea, Aug. 20–22, 2008, pp. 481–486.
IEEE Trans. Aerosp. Electron. Syst., vol. 53, no. 1, pp. 2–11, [34] A. Mahmood, J. W. Wallace, and M. A. Jensen
Feb. 2017. Radio frequency UAV attitude estimation using direction of
[24] S. D. Silverstein, J. M. Ashe, G. M. Kautz, F. W. Wheeler, and arrival and polarization
A. Jacomb-Hood In Proc. Eur. Conf. Antennas Propag., Paris, France,
Tripulse: A system for determining orientation and attitude of Mar. 19–24, 2017, pp. 1857–1859.
a satellite borne active phased array [35] A. Mahmood, J. W. Wallace, and M. A. Jensen
IEEE Trans. Aerosp. Electron. Syst., vol. 38, no. 1, pp. 2–12, UAV attitude estimation using antenna arrays
Jan. 2002. In Proc. IEEE Antennas Propag. Soc. Intl. Symp., San Diego,
[25] M. A. M. Marinho et al. USA, Jul. 9–14, 2017, pp. 1863–1864.
Antenna array based positioning scheme for unmanned aerial [36] P. H. Zipfel
vehicles Frames and coordinate systems
In Proc. ITG Workshop Smart Antennas, Stuttgart, Germany, In Modeling and Simulation of Aerospace Vehicle Dynamics,
Mar. 13–14, 2013, pp. 1–6. Reston, VA, USA: Amer. Inst. Aeronaut. Astronaut., 2007.
[26] J. S. Knogl, P. Henkel, and C. Günther [37] L. C. Godara
Attitude estimation based on multibeam antenna signal power Application of antenna arrays to mobile communications. II.
levels Beam-forming and direction-of-arrival considerations
In Proc. ELMAR Symp., Zadar, Croatia, Sep. 25–27, 2013, In Proc. IEEE, vol. 85, no. 8, pp. 1195–1245, Aug. 1997.
pp. 341–344. [38] H. Krim and M. Viberg
[27] M. Weiqing, R. Liu, Y. Xinxin, and E. Kamel Two decades of array signal processing research: The paramet-
Analysis of precision estimation of RF metrology in satellite ric approach
formation flying IEEE Signal Process. Mag., vol. 13, no. 4, pp. 67–94, Jul. 1996.
In Proc. Intl. Conf. Wireless Commun. Signal Process., Nanjing, [39] A. Taflove and S. C. Hagness
China, Oct. 15–17, 2015, pp. 1–5. Computational Electrodynamics: The Finite Difference Time
[28] J. P. C. L. da Costa, S. Schwarz, and L. F. de A. Gadêlha Domain Method. Norwood, MA, USA: Artech House, 2000.
Attitude determination for unmanned aerial vehicles via an [40] J. W. Wallace
antenna array Modeling electromagnetic wave propagation in electrically
In Proc. ITG Workshop Smart Antennas, Dresden, Germany, large structures
Mar. 7–8, 2012, pp. 264–268. Ph.D. dissertation, Provo, UT, USA: Brigham Young Univ.,
[29] T. F. K. Cordeiro, J. P. L. da Costa, K. Liu, and G. A. Borges 2002.
Kalman-based attitude estimation for an UAV via an antenna
array
In Proc. Intl. Conf. Signal Process. Commun. Syst., Gold Coast,
Australia, Dec. 15–17, 2014, pp. 1–10.

Jon W. Wallace (S’99–M’03–SM’13) received the B.S. (summa cum laude) and Ph.D.
degrees in electrical engineering from Brigham Young University, Provo, UT, USA, in
1997 and 2002, respectively.
He was with Jacobs University, Bremen, Germany, from 2006 to 2013, and Wavetronix,
LLC, Provo, UT, USA, from 2013 to 2014. Since 2014, he has been with the Department
of Electrical and Computer Engineering, Lafayette College, Easton, PA, USA. His current
research interests include physical layer security, MIMO communications and radar, and
unmanned aircraft systems.
Dr. Wallace was a recipient of the National Science Foundation Graduate Fellowship
in 1998. In 2002, he received the Harold A. Wheeler Applications Prize Paper Award
in IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION. He was an Associate Editor
for the IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION, IEEE TRANSACTIONS ON
WIRELESS COMMUNICATIONS, and IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC
SYSTEMS.

WALLACE ET AL.: COOPERATIVE RELATIVE UAV ATTITUDE ESTIMATION USING DOA AND RF POLARIZATION 2699

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.
 Attiya Mahmood (S’13–M’19) received the B.Sc. degree in computer engineering from
the University of Engineering and Technology, Taxila, Pakistan, in 2007, the M.Sc.
degree in computer engineering from the National University of Science and Technology,
Rawalpindi, Pakistan, in 2009, and the Ph.D. degree in electrical and computer engineering
from Brigham Young University, Provo, UT, USA, in 2018.
From 2009 to 2011, she worked as a Lecturer with the Institute of Space Technology,
Pakistan, and later joined Jacobs University, Germany as a Research Associate until
January 2013. She is currently working as a Research Engineer with Wavetronix LLC,
Provo, UT, USA. Her research interests include physical layer security, radars, MIMO
communications, and array signal processing.

Michael A. Jensen (S’93–M’95–SM’01–F’08) received the B.S. and M.S. degrees from
Brigham Young University (BYU), Provo, UT, USA, in 1990 and 1991, respectively, and
the Ph.D. degree from the University of California, Oakland, LA, USA, in 1994, all in
electrical engineering.
Since 1994, he has been with the Department of Electrical and Computer Engineering,
BYU, where he is currently a University Professor and Dean of the Ira A. Fulton College of
Engineering. His research interests include antennas and propagation for communications,
microwave circuit design, multiantenna signal processing, and physical layer security.
Dr. Jensen is Past President of the IEEE Antennas and Propagation Society. He was
previously the Editor-in-Chief of the IEEE TRANSACTIONS ON ANTENNAS AND PROPA-
GATION as well as an Associate Editor for the same journal and for the IEEE ANTENNAS
AND WIRELESS PROPAGATION LETTERS. He is a member and Chair of the Joint Meetings
Committee for the IEEE Antennas and Propagation Society, a member of the society
AdCom, a member of the society Publications Committee, and Co-Chair or Technical
Program Chair for six society-sponsored symposia. In 2002, he received the Harold A.
Wheeler Applications Prize Paper Award in the IEEE TRANSACTIONS ON ANTENNAS AND
PROPAGATION. He was elevated to the grade of IEEE Fellow in 2008 in recognition of his
research on multiantenna communication.

Rashid Mehmood (S’05–M’16) received the B.Sc. (cum laude) degree in communication
systems engineering from the Institute of Space Technology, Islamabad, Pakistan, in 2007,
the M.Sc. in electrical engineering from Jacobs University Bremen, Bremen, Germany, in
2010, and the Ph.D. degree in electrical and computer engineering from Brigham Young
University, Provo, UT, USA, in 2015.
He is currently with the radar systems research group, Wavetronix, Provo, UT, USA.
His current research interests include over-the-air testing, reconfigurable antennas, opti-
mization techniques, MIMO communications and radar, and physical layer security.
Dr. Mehmood was a recipient of the IEEE AP-S Undergraduate Research Award, in
2009. In 2012, he was awarded the Brigham Young University High Impact Doctoral
Research Assistantship Award.

2700 IEEE TRANSACTIONS ON AEROSPACE AND ELECTRONIC SYSTEMS VOL. 56, NO. 4 AUGUST 2020

Authorized licensed use limited to: Zakir Husain College of Engineering & Technology. Downloaded on July 10,2021 at 11:31:29 UTC from IEEE Xplore. Restrictions apply.