1

Unmanned Aerial Vehicle Relative Navigation in

GPS Denied Environments
Jeremy Hardy, Jared Strader, Jason N. Gross, Yu Gu West Virginia University
Mark Keck, Joel Douglas, Systems & Technology Research
Clark N.Taylor , Air Force Research Laboratory

Abstract—This paper considers the problem of target handoff B. Previous Works

between Unmanned Aerial Vehicles (UAVs) in a GPS denied
environment, and focuses on the design and evaluation of an Recent work involving relative or cooperative navigation
estimation strategy for determining the relative pose of the assuming GPS denied environments successfully make use of
aircraft. The estimation approach presented in this paper has communicating relative heading and range updates between
three distinct components that act in concert to achieve the platforms in order to counteract IMU drift. For example,
overall objective. First, a novel cooperative control and estimation
strategy is used to determine relative pose from IMU and peer-to- Sharma and Taylor [2] implement the concept in a simplified
peer ranging radio data without any a priori knowledge of either two dimensional simulation of several Miniature Air Vehicles
aircraft’s pose. Next, a relative pose measurement is calculated (MAVs) and report improvements in pose error compared to
using extracted features from downward looking cameras on the IMU only estimation. Additionally, Knuth and Barooah further
two UAVs. The computer vision technique first uses an indexing demonstrate the advantages in [3] by conducting a simulation
scheme based on a hierarchical statistical model to determine
which frames from the two cameras have overlapping coverage, in which teams of mobile platforms communicate with one
aligns the overlapping frames, and then calculates the relative another using IMU and relative pose data to estimate pose
pose estimate. Finally, a nonlinear Kalman Filter, which has been with respect to an assigned starting point. Knuth and Barooah
initialized with the a priori solution from the initialization filter also confirm this improvement in an experimental setting in
is used to estimate relative pose by predicting it through the which two rovers use optical methods to develop relative pose
integration of IMU data of both UAVs with measurement updates
from peer-to-peer radio ranging radios, magnetometers and the measurements. Although both of the referenced works above
computer vision estimates. take advantage of information provided by other surrounding
platforms to improve pose estimation, both of these examples
Index Terms—GPS-Denied Navigation, Relative Inertial Navi-
gation, Relative Pose Computer Vision estimate the pose of all involved vehicles with respect to a
common, stationary origin. In contrast, this paper explicitly
estimates the pose of one aircraft with respect to the body
I. I NTRODUCTION frame of another aircraft.
The remainder of this paper is organized as follows. Sec-
A. Background and Motivation tion II details the formulation of algorithm for relative pose
CCURATE relative navigation of groups of UAVs is de- estimation developed for this study. Section III outlines the
A sirable for many applications involving surveillance, dis-
aster response, formation flight and rendezvous applications,
process used in the simulation environment and presents the
characteristics of the Monte Carlo setup. Section IV presents
such as aerial re-fueling. Many of these applications require the results of the Monte Carlo analysis. Finally, Section V
maintaining accurate navigation in environments in which GPS offers conclusions and discusses plan for future continued
is not always available, either due to malicious GPS jamming work on this topic.
and spoofing [1] or due to natural signal blockages from urban
canyons, dense foliage, caves, or tunnels. For this work, our II. A LGORITHM F ORMULATION
specific motivation is to enable target handoff between two
UAVs flying in a GPS-denied environment. As such, this paper Figure 1 shows the assumed sensor payload on the two
details and assesses the performance of a relative navigation UAVs simulated for testing in this study, where platform A is
algorithm that estimates the position and attitude of a UAV assumed to be the UAV currently tracking a target of interest
with respect to the body-frame of another UAV without the and platform B is having the target handed off to it. Both
use of GPS. platforms are assumed to have a downward facing camera,
The goal of this paper is to characterize the sensitivity of the a peer-to-peer ranging radio system, and tri-axial IMU and
proposed relative pose estimation approach when considering magnetometer sensors. A communication link for transmitting
different sensor characteristics, sensor update frequency, initial IMU and magnetometer data and information for the computer
condition errors and flight dynamics. Through a Monte Carlo vision updates is also assumed. The diagram also displays
analysis, the most important filter design parameters are iden- the assumptions of negligible differences in the magnetic and
tified. This will support a candidate design for future hardware gravitational fields between platforms, which are leveraged in
implementation and flight testing. the filter formulation.
 2

Fig. 1. UAV Platform Payload Diagram and Assumptions

A. Relative Position Initialization Filter 1) If the platforms are traveling along intersecting lines,
For complete details of this filter’s formulation and a there are exactly four solutions for the relative position.
performance analysis, the reader is pointed to a companion 2) If the platforms are traveling parallel with different
paper [4]. Here, we motivate the high-level concept behind its velocities, there are exactly two solutions for the relative
design. position.
In this method, there is no prior information needed about 3) If the platforms are traveling parallel with equal veloc-
the relative pose of each platform. Instead, the motion of ities, there are an infinite number of solutions for the
each platform is used to construct a graph with the range relative position.
measurements between platforms and principal component 4) If the platforms are traveling along the same line in
distance traveled over multiple locations. The constructed opposite directions, there is exactly one solution for the
graph presented in Figure 2 is analogous to a five bar linkage. relative position.
In this problem, only case 1 and case 2 are considered
because case 3 and case 4 are easily detectable and are not
as likely as case 1 and case 2. Using the constructed linkage,
the relative position in polar coordinates can be obtained by
solving the loop closure equations of the linkage.

α cos θ3 = β (1)

a tan2 (θ2 /2) + b tan (θ2 /2) + c = 0 (2)

In Equation 1, α and β are constant and are a function of

the link lengths. In Equation 2, a, b, and c are constant and
are a function of the θ3 and the link lengths. If the system
is free of noise, Equation 1 and Equation 2 can simply be
solved substituting the appropriate values for the ranging radio
Fig. 2. Trajectories of platforms A and B for three locations where r1 , r2 ,
r3 , r4 , and r5 are the link lengths
measurements and the principal component distance. Since the
measurements are noisy, linear least squares is used to solve
The distance measurements obtained from the ranging for the cosine of the relative orientation since this value is
radios and the principal component distance traveled over constant over any number of time steps. Using the least squares
multiple locations form the links, and the platform position estimate of the cosine of the relative orientation, each of the
at each location form the joints. To form this linkage, the four possible solutions for the relative position are estimated
platforms are required to travel at constant velocities. As listed in polar coordinates using nonlinear least squares to solve
below, there are four possible scenarios for the solution of Equation 2. Note the value being estimated in Equation 2
this problem while the platforms are traveling at constant is the angular coordinate of the relative position. The radial
velocities. For each of these cases, there is exactly one solution coordinate is obtained from the ranging radio measurements.
for the relative orientation, but the singularity and the number Since this formulation is sensitive to noise, the solutions
of solutions of the relative position varies depending on the are smoothed using exponential smoothing, then an extended
path of the platforms. Kalman Filter (EKF) is formed for each solution using INS
 3

for predicting the relative pose and the smoothed angular Lastly, given the set of correspondences between any two
coordinate for updating the measurements. Once the filters images and the 3×3 intrinsic camera parameter matrix, K, the
converge, the coordinate frame transformation is calculated for five-point algorithm [7] can be used to compute an estimate
each of the four solutions. At this point, one of the platforms of the essential matrix.
must turn in a different direction and continue traveling at a
B B
new constant velocity. Using the estimated coordinate frame E = RA [tA ] (6)
transformation, the relative position is propagated during this B
Where RA represents the relative rotation from platform A’s
maneuver using INS. Now, the expected distance between
coordinate frame to platform B’s and tB A represents the relative
platforms can be calculated using the INS solution. Using
translation between the two coordinate frames. Thus, a 3D
the difference between the ranging radio measurement and the
point from platform A’s coordinate frame can be transformed
expected distance, the likelihood is calculated for each of the
into platform B’s coordinate frame given the true relative pose.
four solutions. The solution with the highest likelihood is the
solution nearest to the correct solution.
xB = RA
B A
x + tB
A (7)
B. Computer Vision Relative Pose Estimation Since E is being recovered from correspondences, only the
A commonly used model of image formation is the pinhole direction of tB
A can be computed due to the projective ambi-
camera model, which assumes that the model for a camera p guity.
can be parameterized using 3 elements: a 3×3 upper triangular The relative pose from those correspondences is calculated
p
intrinsic camera parameter, K p , a 3 × 3 rotation matrix, RW , using the five-point algorithm in a RANdom Sampling and
representing a rotation of 3D world points to camera-centered Consensus (RANSAC) [8] loop for robustness.
3D points, and lastly a 3 element vector representing camera
p’s position in world coordinates, cp . Given these parameters C. Relative Navigation Filter
the full 3 × 4 perspective projective matrix can be defined by The nonlinear estimator used is an unscented Kalman filter
Equation 3. (UKF). For details on UKF implementations, the readers are
p p p
referred to [9] and [10]. The states estimated by the filter
P p = K p [RW − RW c ] (3) are relative position, relative velocity, and relative orientation
Then the transformation projecting a 3D world point y onto between platform B and platform A. All of the estimated states
2D image point xp using the homogeneous coordinate repre- are represented in the body frame of platform B.
sentation can be defined by Equation 4.  B 
xAB
 B 
x̃p = P p ỹ (4)  yAB 
 B 
z 
Supposing that two such camera models for cameras A and B  AB  
 B 
rB

were available the relative pose can be directly calculated by  ẋAB 
 B   AB B 
subtracting the positions and calculating the rotation matrix x=  ẏAB  = vAB  (8)

between camera A and B.  B 
 żAB  B
θAB
 B 
B B A T
α 
RA = RW (RW ) (5)  AB 
 B 
 βAB 
With the goal of associating pixels between two images and B
γAB
calculating the relative transformation a three step process is
employed. First, the set of images is searched from the two Where xB AB represents the relative position of platform A
cameras to find corresponding images using a modification of with respect to platform B in the x-axis of the body frame
B
the latent Dirichlet allocation (LDA) [5] coupled to a Hidden of platform B. Similarly, αAB represents the relative rotation
Markov Model (HMM). We extract features from the imagery of platform A around the x-axis of platform B’s body axis
and perform vector quantization in feature space to create a in order to match orientation with platform B. The nonlinear
discrete set of ”words” (so named because the LDA model Kalman filter predicts relative pose using relative INS integra-
was developed for document analysis). The LDA model is a tion and issues measurement updates provided by computer
hierarchical statistical model that captures the co-occurrence vision, peer-to-peer ranging radio, and magnetometer sensor
of features in imagery. Finding corresponding images proceeds data. The filter estimates the relative position and velocity
by using the statistical model to calculate the maximum vectors of platform A in the body-frame of platform B, as
likelihood match of the set of features from platform B to well as the attitude that represents the transformation from
B
the models learned on the images of platform A. The HMM platform A to B, RA .
is used to enforce time consistency between observations. 1) Relative Inertial Navigation: The inertial navigation
Next, we use a frame-to-frame matching to align the top formulation used in this study was also used by Frosbury and
matching candidate images from the set of corresponding im- Crassidis in [11], where the authors derived full navigation
ages. A feature-matching approach using the Scale Invariance equations and error-state expressions for relative navigation
Feature Transform (SIFT) [6] is used in this study. applications. Frosbury and Crassidis presented this work for its
 4

implications in formation flight control systems and conducted q

B,+
a simulation to prove the validity of the derived equations. In ||rAB ||2 = (xB,+ 2 B,+ 2 B,+ 2
AB ) + (yAB ) + (zAB ) (14)
this paper, we slightly deviate from [11] by assuming that there
is no gradient in the gravity field between the two platforms. The magnetometer observation model assumes the same
For predicting relative attitude, a relative quaternion is magnetic field vector is acting on each aircraft due to their
integrated using data from both platforms IMUs based on state assumed proximity. Using this assumption the relative attitude
transition relationship derived in [12]. between the platforms is used to rotate body axis magnetome-
ter data from platform A into platform B’s body frame. Then,
+
qAB −
= [Ω(ωB )Γ(ωA )]qAB (9) the actual measurement from the magnetometer sensors on
platform B can be compared to the estimation made from the
Where Ω(ωB ) are Γ(ωA ) are given by: equation below.
" #
−[ωBˆ] ωB B
MB = R A MA (15)
Ω(ωB ) = T
(10)
−ωB 0
The computer vision technique described in Section II-B
and directly provides relative pose estimates in the form of a unit
" # vector from platform B to platform A, as well as a relative
[ωAˆ] −ωA orientation measurement. The measurement error-covariance
Γ(ωA ) = T
(11) matrix, R, in the Kalman filter is a diagonal matrix made
ωA 0
up of the squares of assigned error values used in the sensor
Where ˆ indicates the skew-symmetric matrix of the rotation simulation. For magnetometers and ranging radios this is
vector. Using the updated attitude the relative velocity between straightforward. The covariance for the computer vision mea-
platforms A and B must take into account the Coriolis accel- surements was generated by bounding the empirical statistical
eration terms due to the rotating body frame of platform B. error model derived from image data with a Gaussian that
included 99% of the errors.
B,+ B,− B,− B,−
vAB = vAB − (ω˙B × rAB )τs − ([ωBˆ][ωBˆ]rAB )τs −
B,− B
(12) III. A LGORITHM S ENSITIVITY A NALYSIS
2(ωB × vAB )τs + ∆vB − RA ∆vA
To assess performance sensitivity, a Monte Carlo study was
Where ∆vA and ∆vB are the IMU incremental velocity conducted to investigate the effects of varying flight geometry,
changes measured in the body-frame of platforms A and B, sensor accuracy, sensor measurement frequency, and initial
respectively. It can be noted that there is no modeled acceler- condition error. As an example, Figure 3 and 4 display the
ation due to gravity present in this equation. It is assumed that geometry of Flight Path 2. The three flight paths differ in
the two platforms are close enough to consider the gravity term the initial separation between each platform, with Flight Path
in the Earth Centered Inertial (ECI) coordinates to be the same 1 having the largest separation between aircraft and Flight
for both platforms. Finally, the relative position is predicted Path 3 having the smallest. Each of these simulated paths
using the updated relative velocity estimates as follows: is approximately 8 minutes long, during which the aircraft
converge closer to one another to simulate the use of a control
B,+ B,− B,−
rAB = rAB + (vAB )τs (13) system which would navigate platform B to platform A for
target hand-off. In reality, this control would be driven by the
The above equations make up the nonlinear state transition
initialization filter described in [4]. These simulations were
functions f of the filter used in this study. The diagonal process
conducted in an environment developed at WVU and modified
noise covariance matrix, Q, is constructed using IMU param-
for the purposes of this study, more details of the simulation
eters. Velocity Random Walk (VRW) and Angular Random
environment can be found in [14].
Walk (ARW) are used to construct the matrix in the same
The error characteristics of the HG1930-CA50 IMU was
way they would for a single platform INS, √ with the exception
used as a baseline for this study and where scaled to error
that the matrix is multiplied by a scalar 2 to account for
ranges of various grade IMUs [15] and shown in Table I .
the addition of two random error sources (from two separate
The baseline Honeywell sensor has a gyroscope bias in-run
platforms) instead of one.
stability of 1.0 deg
hr , a accelerometer bias in-run stability of
2) Measurement Updates: Measurement updates from f ps
0.3mg, a Velocity Random Walk (VRW) of 0.3 √ and an
peer-to-peer ranging radio, magnetometer, and computer vi- deg
hr
sion methods help to minimize the drift present in the relative Angular Random Walk (ARW) of 0.125 √hr in the roll axis
deg
INS. The observation model for ranging radio measurements is and 0.09 √ hr
in the pitch and yaw axes [16]. Additionally,
the magnitude of the predicted relative position. The estimated the simulation assumed baseline error characteristics for the
B,+
distance, ||rAB ||2 , can be compared to the measured distance ranging radio to be white Gaussian noise σ = 5cm and
provided by ranging radio sensor data. Here we build upon magnetometer measurement errors of white noise Gaussian
recent work that has used Ultra Wideband (UWB) peer-to- with σ = 250, 000nT, respectively. Tables II and III show the
peer radio in a GPS-degraded setting for relative navigation update rates and error characteristics (Error Scale Factor, ESF)
applications for formation flight in [13]. used for the magnetometer, computer vision, and ranging radio
 5

as well as the distribution of initial condition errors used in

the Monte Carlo analysis. TABLE II
M AGNETOMETER , R ANGING R ADIO , AND C OMPUTER V ISION M ONTE
C ARLO C HARACTERISTICS

Sensor Scaling Frequency (Hz) ESF

Magnetometer 100 1-5
Ranging Radio 5, 10, 25, 50, 100 1-5
Computer Vision 1, 2, 5, 10, 25 1

TABLE III
I NITIAL C ONDITION E RROR D ISTRIBUTIONS

Characteristic Mean Std. Dev.

Position (m) 0 5
Velocity ( m
s
) 0 10
Attitude (deg) 0 2

Fig. 3. North v East Representation of Flight Path 2

TABLE IV
IMU Q UALITY M ONTE C ARLO D ISTRIBUTION

IMU Quality # of Runs

IMU 1 126
IMU 2 112
IMU 3 124
IMU 4 138

TABLE V
C OMPUTER V ISION U PDATE F REQUENCY M ONTE C ARLO D ISTRIBUTION

Computer Vision
# of Runs
Frequency (Hz)
1 105
Fig. 4. ENU Representation of Flight Path 2 2 95
5 101
10 111
25 88
TABLE I
I NERTIAL M EASUREMENT U NIT S CALING FACTORS

IMU Characterization ESF

Automotive Grade 50 TABLE VI
Tactical Grade 1 R ANGING R ADIO U PDATE F REQUENCY M ONTE C ARLO D ISTRIBUTION
Intermediate Grade 1/100
Aviation Grade 1/1000 Ranging Radio
# of Runs
Frequency (Hz)
Tables IV - VII show the distribution of different character- 5 89
10 108
istics of the simulation runs included of the 500 Monte Carlo 25 108
trials that were executed. 50 108
100 87

IV. R ESULTS TABLE VII

F LIGHT PATH M ONTE C ARLO D ISTRIBUTION
A. Initialization Filter
Flight Path # of Runs
The performance analysis of the initialization filter are 1 164
described in [4]. In this study, we assume an a prior initial 2 169
condition error for the relative pose Kalman filter as shown in 3 167
Table III.
 6

B. Standalone Computer Vision

In order to evaluate computer vision measurements, a
two-component Gaussian mixture error model of the five-
dimensional relative pose parameter space was estimated. The
model was estimated by comparing rotation and translational
direction from the computer vision pose estimate on a set
of measured data compared to the solution using GPS/INS
output. Using 241 test images the best matching images were
computed using the LDA-HMM model. Then SIFT matching
followed by the five-point algorithm was applied on the query
image and the top matching image to estimate the relative
pose. The RANSAC estimation process required at least 100
inliers to improve robustness, reducing the number of data
points to 159. Fig. 6. Error in Translation Vector Estimate from Five-Point Algorithm
The error was computed in both the estimated rotation
angles and the difference between the estimated translation
vector and the true translation vector in 3 dimensions. An in the computer vision system’s direction of translation update
example of the error in rotation angles in shown in Figure 5. are projected over a long baseline and the relative dynamics
Each of the 3 plots shows error in a different angle, computed between the UAVs are more exaggerated. As such, our analysis
with respect to the ground-truth axis-angle representation; thus primarily considers the performance over the last 2/3 of each
the angles do not directly correspond to roll, pitch, and yaw, flight. For reference, Figure 7 compares the median error of
but with the error in local coordinates. For the majority of the the 500 Monte Carlo trials over the last 2/3 of each flight
sequence the error is within a degree, with some earlier parts where the median is 12 m 3D RSS for position error and 0.5
of the sequence having a larger error. The beginning portion 3D RSS for attitude error.
of the sequence is less feature-rich, and therefore, the five-
Overall Performance N=500 trials
point algorithm is not quite as robust as later in the sequence. 1
Similarly, in Figure 6, the error in translation vector estimate is 0.8
shown. Note here the clear mode at the origin, with erroneous 0.6
translation vectors also scattered about space. 0.4
0.2
0
10 1 10 2 10 4 10 6 10 8
3D RSS Error (m)
1
0.8
0.6
0.4
0.2
0
10 -1 10 0 10 1 10 2
3D RSS Attitude (deg)

Fig. 7. CDF of Position and Attitude Error from 500 trials (considering last
2/3 of each flight)

Fig. 5. Error in Rotation Estimate from Five-Point Algorithm

All the remaining figures show distributions based upon
From this data a two-component 6D Gaussian error model last 2/3 of each simulated flight separated by run various
for computer vision-based measurements of relative pose was characteristics. Figures 8 - 14 display Cumulative Distribution
inferred. This model is used in the simulation of computer Functions (CDFs) for specified simulation characteristics. The
vision measurements and the variance of the error is used in position and attitude errors displayed in these plots are the
the relative pose filter. two norm of the Root Mean Square (RMS) error. These plots
show the performance of the relative pose filter during the
Monte Carlo analysis for a specific simulation characteristic,
C. Relative Navigation Filter while all other values of the simulation remain random. For
For each flight path simulated, the UAVs start far away from example, Figure 9 displays the results of each filter run sorted
one another and eventually converge to a similar trajectory. by the quality of IMU used during the simulations, while all
Over this profile, the most important period for target handoff other Monte Carlo variables are randomized. This allows for
is after the UAVs have converged to a similar trajectory. During the evaluation of the effect IMU quality has on the filter’s
the initial period of the flight when the UAVs are far apart, the ability to estimate relative pose.
estimation error of the filter is relatively poor because errors Figure 8 shows a change in filter accuracy based on flight
 7

geometry, where the closer the starting positions of the aircraft Sensitivity to Computer Vision Update Rate
1
the more accurate the pose estimation than the two with larger
0.8
initial seperation. 0.6
0.4
Sensitivity to Flight Path 0.2
1
0
0.8 1 2 4 6 8
10 10 10 10 10
0.6 3D RSS Relative Position (m)
0.4
1
0.2
0.8
0 1 Hz
10
1
10
2
10
4
10
6
10
8 0.6 2 Hz
3D RSS Relative Position (m) 0.4 5 Hz
0.2 10 Hz
1 25 Hz
0
0.8
10 -1 10 0 10 1 10 2
0.6 3D RSS Relative Attitude (deg)
Path 1
0.4 Path 2
0.2 Path 3
0 Fig. 10. CDF of Position and Attitude Error from Varying Computer Vision
10 -1 10 0
10 1
10 2 Update Frequency (considering last 2/3 of each flight)
3D RSS Relative Attitude (deg)

sensitivity on the filter’s estimation performance when varying

Fig. 8. CDF of Position and Attitude Error from Varying Flight Geometry
(considering last 2/3 of each flight) ranging radio update rate, ranging radio and magnetometer
error magnitudes, and initial condition errors. That is, each of
We believe that this discrepancy is likely attributed to the the CDFs are quite similar for each remaining case.
flights with smaller initial separation containing fewer and less
dramatic dynamic maneuvers coupled with the fact that as the Sensitivity to Ranging Radio Update Rate
separation of the UAVs increases, the errors in the computer- 1
0.8
vision based estimates of the direction of translation result in
0.6
a larger cartesian error. 0.4
Figure 9 shows that IMU grade is not very critical for 0.2
positioning accuracy (i.e. all CDFs are similar), though, as ex- 0
pected, the automotive grade IMUs shows noticeable attitude 10 1 10 2 10 4 10 6 10 8
estimation performance degradation. 3D RSS Relative Position (m)
1
Sensitivity to IMU Quality 0.8
5 Hz
1 0.6 10 Hz
0.8 0.4 25 Hz
0.6 0.2 50 Hz
0.4 0 100 Hz
0.2 10 -1 10 0 10 1 10 2
0 3D RSS Relative Attitude (deg)
10 1 10 2 10 4 10 6 10 8
3D RSS Error (m) Fig. 11. CDF of Position and Attitude Error from Varying Ranging Radio
1 Update Frequency (considering last 2/3 of each flight)
0.8
IMU 1
0.6 IMU 2 The fact that the filter formulation produces reasonably
0.4 IMU 3 accurate position and attitude estimates when considering the
0.2 IMU 4
ranges of sensor quality and availability tested as shown in
0
10 -1 10 0 10 1 10 2 Figure 7, is an important determination as we move toward
3D RSS Attitude (deg) filter and payload design for flight demonstrations. For a
baseline design if we only consider trials that use the HG-
Fig. 9. CDF of Position and Attitude Error from Varying IMU Quality 1930 IMU tactical grade IMU, 5 Hz computer vision updates
(considering last 2/3 of each flight) and assume a favorable flight path, such as flight path 3, the
However, the most pronounced differences in estimation median 3D RSS position error of 13 Monte Carlo trials 21
accuracy is a result of the computer vision update frequency. meters and the median 3D RSS attitude error is 2.7 degrees.
Figure 10 displays an obvious trend, in which the filter
provides less accurate pose estimations at lower computer V. C ONCLUSIONS AND F UTURE W ORK
vision frequency updates, indicating that these updates are This study presents a relative pose estimation algorithm
crucial to the filters ability to combat IMU drift. and show its sensitivity to flight geometry, sensor character-
Interestingly, the remainder of the sensitivities considered as istics and initial condition errors. The algorithm is a GPS-
shown in Figures 11 - 14 indicate that there is relatively little denied relative navigation filter that primarily consists of an
 8

Sensitivity to RR & Mag. ESF

1
0.8
0.6
0.4
0.2 0 to 1
1 to 2
0
1 2 4 6 8 2 to 3
10 10 10 10 10 3 to 4
3D RSS Relative Positon (m) 4 to 5
1
0.8
0.6
0.4
0.2
0
10 -1 10 0 10 1 10 2
3D RSS Relative Attitude (deg)

Fig. 12. CDF of Position and Attitude Error from ESF Values of Ranging Radio and Magnetometer Sensors (considering last 2/3 of each flight)

Sensitivity to Initial Pos./Vel. Error

1
0.8
0.6 0 to 5m
5 to 10m
0.4 10 to 15m
0.2 15 to 25m
0
1 2 4 6 8
10 10 10 10 10
3D RSS Relative Positon Error (m)
1
0.8
0.6
0 to 5m/s
0.4 5 to 10m/s
0.2 10 to 15m/s
0
10 1 10 2 10 4 10 6 10 8
3D RSS Relative Position Error (m)

Fig. 13. CDF of Position Error from Initial Condition Errors in Position and Velocity Estimation (considering last 2/3 of each flight)

Sensitivity to Initial Attitude Error

1
0.8
0.6
0.4
0.2
0
10 1 10 2 10 4 10 6 10 8
3D RSS Relative Positon Error (m)
1
0.8
0.6 0 to 1 deg.
1 to 2 deg.
0.4 2 to 3 deg.
0.2 3 to 4 deg.
0
10 -1 10 0 10 1 10 2
3D RSS Relative Attitude (deg)

Fig. 14. CDF of Position and Attitude Error from Initial Condition Errors in Attitude Estimation (considering last 2/3 of each flight)

integration of inertial navigation and computer vision. For and used to assess updates of this nature being fused with
computer vision updates, a statistical model was used to other sensors. Based on this analysis, it is shown that this
rapidly determine overlapping content between two sets of algorithm can provide relative pose estimates in a GPS-
imagery such that relative pose can be extracted. Using a set denied setting with enough accuracy to assist in the problem
of flight data imagery, a statical error model was developed of platform target handoff. In particular, considering overall
 9

sensitivity to sensor quality and availability, simulated trials

were shown to have 10-meter-level position accuracy and 0.5
degree accuracy in a median sense. Our Monte Carlo analysis
also shows that the update rate of computer vision relative
pose measurements, quality of IMU, and the separation of
flight paths have the most prominent impact of performance.
Furthermore, the incorporation of magnetometer updates, the
quality of ranging radio measurements, and initial condition
errors are shown to not have a significant impact on the overall
algorithm performance. Future work will include end-to-end
system development for experimental testing.

ACKNOWLEDGEMENT
This material is based upon work supported by the United
States Air Force under Contract No. FA8650-15-M-1943.
Approved for public release, case number 88ABW-2016-0102.

R EFERENCES
[1] A. J. Kerns, D. P. Shepard, J. A. Bhatti, and T. E. Humphreys,
“Unmanned aircraft capture and control via gps spoofing,” Journal of
Field Robotics, vol. 31, no. 4, pp. 617–636, 2014.
[2] R. Sharma and C. Taylor, “Cooperative navigation of mavs in gps denied
areas,” in Multisensor Fusion and Integration for Intelligent Systems.
IEEE, 2008.
[3] J. Knuth and P. Barooah, “Collaborative 3d localization of robots from
relative pose measurements using gradient descent on manifolds,” in
Robotics and Automation. IEEE, 2012.
[4] J. Strader, Y. Gu, J. Gross, M. DePetrillo, and J. Hardy, “Cooperative
relative localization for moving uavs with single link range
measurements,” in Proceedings of IEEE/ION Position Location and
Navigation Symposium. IEEE/ION, 2016.
[5] D. M. Blei, A. Y. Ng, and M. I. Jordan, “Latent dirichlet allocation,”
Jornal of Machine Learning Research, vol. 3, pp. 993–1022, 2003.
[6] D. Lowe, “Distinctive image features from scale-invariant keypoints,”
International Journal of Computer Vision, vol. 60, no. 2, pp. 91–110,
2004.
[7] H. Stewenius, C. Engels, and D. Nister, “Recent developments on direct
relative orientation,” ISPRS Journal of Photogrammetry and Remote
Sensing, vol. 60, no. 4, pp. 284–294, 2006.
[8] M. Fischler and R. Bolles, “Random sample consensus: A paradigm
for model fitting with application to image analysis and automated
cartography,” Communications of the ACM, vol. 24, no. 6, pp. 381–395,
1981.
[9] M. Rhudy and Y. Gu, “Understanding nonlinear kalman filters, part ii:
An implementation guide,” 2013.
[10] D. Simon, Optimal state estimation: Kalman, H infinity, and nonlinear
approaches. John Wiley & Sons, 2006.
[11] A. M. Fosbury and J. L. Crassidis, “Relative navigation of air vehicles,”
Journal of Guidance, Control, and Dynamics, vol. 31, no. 4, pp. 284–
834, 2008.
[12] R. A. Mayo, “Relative quaternion state transition relation,” Journal of
Guidance, Control, and Dynamics, vol. 2, no. 1, pp. 44–48, 1979.
[13] J. N. Gross, Y. Gu, and M. B. Rhudy, “Robust uav relative navigation
with dgps, ins, and peer-to-peer radio ranging,” Automation Science and
Engineering, IEEE Transactions, vol. 12, no. 3, pp. 935–944, 2015.
[14] R. Watson, V. Sivaneri, and J. Gross, “Performance Characterization
of Tightly-Coupled GNSS Precise Point Positioning Inertial Navigation
within a Simulation Environment,” in 2016 AIAA Guidance Navigation
and Control Conference. AIAA, 2016.
[15] P. D. Groves, Principles of GNSS, inertial, and multisensor integrated
navigation systems. Artech House, 2013.
[16] “Hg1930 datasheet,” https://aerospace.honeywell.com/products/
communication-nav-and-surveillance/inertial-navigation/
inertial-measurement-units/hg1930, 2015.