Two Years of Visual Odometry on the Mars

Exploration Rovers

Mark Maimone, Yang Cheng, and Larry Matthies

Jet Propulsion Laboratory
California Institute of Technology
Pasadena, CA USA
[email protected]

Abstract

NASAs two Mars Exploration Rovers (MER) have successfully demonstrated

a robotic Visual Odometry capability on another world for the first time. This
provides each rover with accurate knowledge of its position, which allows it
to autonomously detect and compensate for any unforeseen slip encountered
during a drive. It has enabled the rovers to drive safely and more effectively in
highly-sloped and sandy terrains, and has resulted in increased mission science
return by reducing the number of days required to drive into interesting areas.
The MER Visual Odometry system comprises onboard software for comparing
stereo pairs taken by the pointable mast-mounted 45 degree FOV Naviga-
tion cameras (NAVCAMs). The system computes an update to the 6 Degree
Of Freedom rover pose (x, y, z, roll, pitch, yaw) by tracking the motion of
autonomously-selected terrain features between two pairs of 256x256 stereo
images. It has demonstrated good performance with high rates of successful
convergence (97% on Spirit, 95% on Opportunity), successfully detected slip
ratios as high as 125%, and measured changes as small as 2 mm, even while
driving on slopes as high as 31 degrees.
During the first two years of operations, Visual Odometry evolved from an
extra credit capability into a critical vehicle safety system. In this paper
we describe our Visual Odometry algorithm, discuss several driving strategies
that rely on it (including Slip Checks, Keep-out Zones, and Wheel Dragging),
and summarize its results from the first two years of operations on Mars.

1 Background

Keeping track of a vehicles location is one of the most challenging aspects of planetary rover
operations. NASAs Mars Exploration Rovers (MERs) typically have been commanded only
once per Martian solar day (or sol) using a pre-scheduled sequence of precise metrically
specified commands (e.g., drive forward 2.34 meters, turn in place 0.3567 radians to the
right, drive to location X,Y, take color pictures of the terrain at location X,Y,Z (Biesiadecki
et al., 2005)), so having an accurate position estimate onboard during the execution of all
terrain-based commands has been of critical importance.
The design goal for MER was to maintain a position estimate that drifted no more than
10% during a 100 meter drive. MER rover onboard position and attitude estimates were
updated at 8 Hz nearly every time the wheels or rover arm (Instrument Deployment Device,
or IDD) were actuated. Changes in attitude (roll, pitch, yaw) were measured using a Litton
LN-200 Inertial Measurement Unit (IMU) that has 3-axis accelerometers and 3-axis angular
rate sensors, and changes in position were estimated by combining attitude measurements
with encoder readings of how much the wheels turned (wheel odometry). Position estimates
derived solely from those sensors easily achieved the desired accuracy in benign terrains (Li
et al., 2005), but not on steep slopes or sandy terrain.

After moving a small amount on a slippery surface, the rovers were often commanded to
use camera-based Visual Odometry to correct any errors in the initial wheel odometry-based
estimate that occur when the wheels lose traction on large rocks and steep slopes. Our
Visual Odometry system computes an update to the 6-DOF rover pose (x, y, z, roll, pitch,
yaw) by tracking the motion of interesting terrain features between two pairs of stereo
images in both 2D pixel coordinates and 3D world coordinates. A maximum likelihood
estimator applied to the computed 3D offsets produces the final motion estimate. However,
if any internal consistency check fails, too few feature points are tracked, or the estimation
fails to converge, then no motion estimate update will be produced and the initial estimate
(nominally based on wheel odometry and the IMU) will be maintained.

NASAs twin Mars Exploration Rovers Spirit and Opportunity landed on the surface of Mars
in January 2004. As shown in the blue lines of the traverse plots in Figures 14 and 15, human
rover drivers have commanded extensive use of the Visual Odometry software, especially
during high-tilt operations: driving Opportunity inside Eagle and Endurance craters, and
climbing Spirit through the Columbia Hills. Visual Odometry was not used on every drive
step, however. Initially the reason was that the operations team was unfamiliar with the
capability; this was an extra credit capability not originally baselined for the mission, and
therefore had not been included in operational readiness tests (although it did go through the
verification and validation program). But even after being shown to work well on Mars, the
time required to perform vision processing on the 20 MHz CPU reduced the overall effective
drive speed by an order of magnitude, so the utility of the better position estimate had to
be weighed against the desire to cover longer distances (Biesiadecki et al., 2007).

In the first two years since landing, the rovers have driven over terrain with as much as 31
degrees of tilt, and over textures comprised of slippery sandy material, hard-packed rocky
material, and mixtures of both. Engineering models of vehicle slip in sandy terrain developed
during Earth-based testing correlated remarkably well with the sand-only terrain inside Eagle
crater during the first two months. However, slip was extremely difficult to predict when the
rover was driven over nonhomogeneous terrains (e.g., climbing over rock for one part of a
drive and loose soil for another, or climbing over low-lying ripples of sandy material). Early
on, the uncertainty in the amount of slip resulting from drives on high slopes or loose soils
forced the operations team to spend several days driving toward some targets, even those
just a few meters away. But through the rest of the mission, Visual Odometry software has
enabled precision drives (i.e., ending with the science target being directly reachable by the
IDD) over distances as long as 8 meters on slopes greater than 20 degrees (see Figure 3),
and made it possible to safely traverse the loose sandy plains of Meridiani.
2 Algorithm

Work on estimating robot motion with stereo cameras can be traced back to Moravecs work
(Moravec, 1980). Following Moravecs work, Matthies et al. treated motion estimation as a
statistical estimation problem and developed sequential methods for estimating the vehicle
motion and updating the landmark models. This system achieved an accuracy of 2% of
distance over 5.5 meters and 55 stereo image pairs (Matthies and Shafer, 1987; Matthies,
1989), with a consistent level of accuracy reported more recently (Olson et al., 2003). Similar
work has been reported elsewhere (Zhang et al., 1988; Lacroix et al., 1999; Nister et al., 2004).
Recently, Nister et al. have reported a successful real-time Visual Odometry implementation
(Nister et al., 2006; Nister, 2004). This implementation contains two motion estimation
schemes: the stereo scheme, which is an iterative pose refinement scheme, and the monocular
scheme, which was based on the framework of 5-point algorithm (Nister, 2004) and an outlier
rejection scheme (Nister, 2005) and good results have been reported on a very long image
sequences. Other approaches to the subject of Visual Odometry schemes have been reported
as well. For example, McCarthy and Barnes have reported the performance of optical flow
based motion estimation (McCarthy and Barnes, 2004) and Vassallo and Gluckman and
others have developed ego-motion estimation with omnidirectional images (Vassallo et al.,
2002; Gluckman and Nayar, 1998; Corke et al., 2004).

Our approach to position estimation is to find features in a stereo image pair and track them
from one frame to the next. The key idea of the present method is to determine the change
in position and attitude for two pairs of stereo images by propagating uncertainty in a 3D
to 3D pose estimation formulation using maximum likelihood estimation. The basic steps of
this method are described as follows.

Figure 1: Feature tracking occurs between every pair of images. In this view, several images
from Spirits Sol 178 drive and their tracked features have been superimposed.

Feature Detection First, features that can be easily matched between stereo pairs and tracked
across a single motion step are selected. An interest operator tuned for corner detection (e.g.
Forstner or Harris) is applied to an image pair, and pixels with the highest interest values are
selected. To reduce the computational cost, a grid with cells smaller than a preset minimum
distance between features is superimposed on the left image. The feature with strongest
corner response in each grid cell is selected as a viable candidate. A fixed number of features
having the highest interest operator responses is selected, subject to a minimum distance
constraint to ensure that features span the image.
Feature-based Stereo Matching Each selected features 3D position is computed by stereo
matching. Because the stereo cameras are well calibrated, the stereo matching is done
strictly along the epipolar line with only a few pixels of offset buffer above and below it. We
use pseudo-normalized correlation to determine the best match. In order to obtain subpixel
accuracy, a biquadratic polynomial is fit to a 3x3 neighborhood of correlation scores, and
the peak of this polynomial is chosen as the correlation peak.

The 3D positions of these selected features are determined by intersecting rays projected
through the camera models. In perfect conditions, the rays of the same feature in the left
and right images would intersect at a point in space. However, due to image noise, camera
model uncertainty and matching error, they do not always intersect. The shortest distance
gap between the two rays indicates the goodness of the stereo match: features with large
gaps are thrown out.

Next we compute the covariance associated with each feature using methods described in
(Matthies and Shafer, 1987) and detailed in (Matthies, 1989).

Note that the covariance of each point P is

" P #
0 l 0
= P P0T (1)
P
P
0
P
r

where P0 is the Jacobian matrix or the first partial derivative of P with respect to the 2D
feature locations in the left and right images, and l and r are 2x2 matrices whose elements
P P

are the curvatures of the biquadratic polynomial along the vertical, horizontal and diagonal
directions, which can be obtained directly from subpixel interpolation.

The quality of a 3D feature is a function of its relative location, the gap between the two
stereo rays, and the sharpness of the correlation peak. This covariance computation fully
reflects these three factors.

Feature Tracking After the rover moves a short distance, a second pair of stereo images
is acquired. The features selected from the previous image are projected into the second
pair using the approximate motion provided by onboard wheel odometry (see Figure 1 for
some examples). Then a correlation-based search reestablishes the 2D positions precisely
in the second image pair. Stereo matching of these tracked features determines their new
3D positions. Because the 3D positions of those tracked features are already known from
the previous step, the stereo matching search range can be greatly reduced. A rigidity test
(comparing the relative 3D distances of corresponding points) is performed as an initial
outlier rejection step.

Robust Motion Estimation If the initial motion is accurate, the difference between two es-
timated 3D feature positions should be within the error ellipse. However, when the initial
motion is off, the difference between the two estimated positions of the 3D points reflects
the error of the initial motion and it can be used to determine the change of rover position.

Motion estimation is done in two steps. First, a less accurate motion is estimated by least
squares estimation. The error residual between current position PCj and previous position
PP j of the j th feature is
 ej = PCj RPP j T (2)

and the cost expression is

wj eTj ej
X
M (R, T) = (3)
X X
1
wj = (det( Pj) + det( Cj )) (4)

There is a closed form solution for this least squares estimation (Schonemann and Carroll,
1970; Matthies, 1989). Let

X
w = wj (5)
j
X
Qc = wj PCj (6)
j
X
Qp = wj P P j (7)
j
X
A = wj PP j PCj (8)
j
1
E = A Qc QTp (9)
w

Let E = USVT be the singular value decomposition of E. Then

R = UVT (10)
1h i
T = Qc RQp (11)
w

The advantage of this method is that it is simple, fast and robust. Its disadvantage is that
its results can be substantially inferior to those derived with a full error model because this
only takes the quality of the observations (the volume of the error ellipsoid or determinant
of the covariance matrix) as a weight factor (Matthies, 1989).

Because it is an inexpensive operation, we embed it within a RANSAC (Random Sample

Consensus) process to do outlier removal:

1. A small set of features (e.g., six) is randomly selected and the motion is then estimated
using the least squares estimation method.
2. All features from the previous step are projected into the current image frame using
the newly estimated motion. If the gap between a reprojected feature and its corre-
spondent is less than a threshold (e.g. 0.5 pixels), the score of this iteration will be
incremented once for each viable feature.
 3. Steps 1 and 2 repeat for a fixed number of iterations and the motion with the highest
score is selected. All features that pass this iteration will be used in the following
more accurate estimation the maximum likelihood motion estimation.

Maximum Likelihood Estimation The maximum likelihood motion estimation considers the
3D feature position difference and associated error models when estimating rover position
(Matthies and Shafer, 1987; Matthies, 1989). As above, let PP j and PCj be the observed
positions of feature j prior to and after the current robot motion. Then

PCj = RPP j + T + ej (12)

where R and T are the rotation and translation of the robot and ej is the combined error
in the observed positions of jth features. In this estimation, three axis rotations R and
translation T are directly determined by minimizing the summation:

eTj Wj ej
X
(13)

where Wj = (R P j RT + Cj )1 is the inverse covariance matrix of ej . The minimization

P P

of this nonlinear problem is done by linearization and an iterative process (Matthies, 1989).
The linearization is obtained by taking a first-order expansion of Equation 12 with respect
to rotation angles 0 (x , y , z ), and applying the least-squares method:

PCj = RPP j + T + ej (14)

R0 PP j + Jj ( 0 ) + T + ej (15)

where the 3x3 Jacobian matrix Jj is given by:

Jj = [Rx PP j Ry PP j Rz PP j ] (16)

Where Rx , Ry , Rz are the partial derivatives of the rotation matrix with respect to the
rotation angles x , y , z , respectively. We obtain a solution for the optimal rotation in a
least-squared sense, using an iterative method:

Pij = PCj Ri1 PP j + Jij1 i1 (17)

1

JTij Wj Pij JTij Wj (

X X X X
i = Wj )1 Wj Pij (18)
j j j j

JTij Wj Jij JTij Wj (

X X X X
Wj )1 Wj Jij
j j j j

Iteration continues until | i i1 | < , where typically = 0.000006. Then the optimal
translation is obtained from:
 Table 1: Optional Constraints on 3D Updates
Constraint Units
Allowed nonconvergences
World Coordinate delta X component meters
World Coordinate delta Y component meters
3D Update Vector Magnitude meters
Max Change in Roll radians
Max Change in Pitch radians
Max Change in Yaw radians
Max Angle from Downslope (when tilt is high) radians
Min Tilt Angle at which to check Downslope radians

X X  
T = ( Wj )1 Wj PCj RPP j (19)
j j

The key advantage of this approach, compared to the scalar-weighted approach (Equation 3),
is that incorporating the full 3D covariance matrices of the 3D features in the matrices W
properly weights the triangulation error, leading to much better motion estimates (Matthies
and Shafer, 1987; Matthies, 1989).

The MER Visual Odometry implementation improves on earlier implementations (Olson

et al., 2003) in two main areas. First, we used a more accurate feature covariance calculation.
In the previous implementation, the feature stereo matching error in image space is a constant
for all features, which cannot faithfully reflect the error model. For example, the correlation
error in an area of high image texture area would be less than the low texture area, which
should be incorporated into the covariance computation. In this MER implementation we
used the curvature of the biquadratic polynomial along the vertical, horizontal and diagonal
directions to quantify the quality of the correlation matching. The second improvement is
the use of RANSAC with the least-squares estimator for outlier rejection based on image
reprojection error of the features, which is similar to the procedure suggested by Nister
(Nister, 2005; Nister et al., 2006).

As of the February 2005 version of MER flight software, optional constraints can also be
placed on the final motion estimate to provide additional sanity checking. The magnitude
of the 3D update vector, its X and Y World Frame components, the magnitude of the roll,
pitch and yaw updates, and the angular deviation from a purely downslope vector can all be
restricted (see Table 1). Any update violating the active set of constraints is treated as an
update failure. Other types of update failures are initialization steps (there is nothing
to track, hence no update) and failures to converge. The total number of acceptable update
failures can also be constrained.

3 Ground-based Validation

Precursor implementations of our Visual Odometry software have been tested on numerous
rover platforms, typically resulting in position estimates accurate to within 2% of the distance
traveled (Matthies and Shafer, 1987; Olson et al., 2003). Details of those tests can be found
Figure 2: Visual Odometry Error measured during a 2.45 meter drive using HAZCAMs
on the MER Surface System Testbed Lite rover. The rover was driven over several large
non-obstacle rocks, each less than 20 cm tall, in 35 cm steps. The vehicle was held in place
during the final step, so the wheel odometry error for that step is artificially large, yet the
Visual Odometry error remains small.

in the cited publications, here we will just summarize a recent test on a research rover and
a quick validation performed on a MER engineering model vehicle.

The latest tests were conducted on JPLs Rocky 8 rover at the JPL Marsyard and in Johnson
Valley, California (Helmick et al., 2004). Rocky 8 has two pairs of hazard avoidance stereo
cameras mounted on the front and rear of the rover body about 50 cm above the ground.
The image resolution is 640 by 480, horizontal and vertical fields of view are 80 degrees
horizontal by 64 degrees vertical and the baseline is about 8.4 cm. The Johnson Valley site
had slopes of loose granular sand where the rover experienced substantial slip, tilt, and roll
during the test.

In order to evaluate Visual Odometry performance, high precision ground-truth data (posi-
tion and attitude) was also collected using a total station (like a surveyors theodolite with
a laser range sensor). By tracking four prisms on top of the rover, the rovers position and
attitude were measured with high precision (< 2 mm in position and < 0.2 degree in atti-
tude). The absolute position errors were less than 2.5% over the 24 meter Marsyard course,
and less than 1.5% over the 29 meter Johnson Valley course. The rotation error was less
than 5.0 degrees in each case (Helmick et al., 2004).

Simple confirmation tests were also run on the MER Surface System Testbed Lite rover in
an indoor sandbox test area. Ground truth was acquired using a total station to measure the
vehicles 6-DOF motion estimate by tracking three points at each step. During these tests
Visual Odometry processing took place using images from the 120-degree FOV HAZCAM
sensors (but on Mars only the 45-degree FOV NAVCAMs are commanded). Several tests
were run in which Visual Odometry was found to be as good as wheel odometry on simple
terrain (within the design spec), and much better in complex terrain.

As an example, Figure 2 shows the position estimation error that resulted from the most
slip-inducing test run on the MER engineering model: a 2.45 meter rock-laden course driven
in 35 cm steps. All 6 rocks that were climbed were small enough (less than 20cm in height)
Figure 3: CGI rendering of Opportunity on the side of Burns Cliff in Endurance Crater.
Opportunity perched here on slopes ranging from 22 to 31 degrees.

that they would not be considered obstacles. The straight sloped line above the light blue
background in the figure represents the design goal of at most 10% error in the position
estimate. The dark curve represents the error that accrued when the position was estimated
using only the IMU and wheel odometry; after 1.4 meters of driving, the accumulated error
had already gone beyond the desired 10% curve. Finally, the light curve at the bottom
represents the error remaining after Visual Odometry processing has completed. Even after
2.45 meters of driving over rough obstacles with as much as 85% slip, the Visual Odometry
error remained small, less than 1% of the total traverse distance.

Other work on quantitative evaluation of a different Visual Odometry algorithm has been re-
ported by Nister et al. (Nister et al., 2006). Their algorithm is formulated to minimize image
reprojection error instead of error between 3D feature coordinates. These authors measure
performance in terms of errors in integrated path length, compared to differential GPS as
ground truth, in order to factor out effects of attitude error. Note that we have used total
position error as the performance measure, which does include the effects of attitude error.
Nister et al. reported errors of 1-2% on data sets covering up to 365 meters. It is difficult to
draw any conclusions from this comparison, since besides using a different formulation and
a different performance metric, this work also had very different computational constraints
that allowed tracking two orders of magnitude more features per frame.

4 Operational Constraints

The requirement that Visual Odometry operate on another world leads to operational con-
straints that make traditional terrestrial solutions inadequate.

For example, consider the feature tracker. Some terrestrial visual odometry systems follow
the update often, search less mantra: they update so frequently that the motion of each
feature between frames is guaranteed to be small (e.g., (Amidi et al., 1999)). This reduces
the amount of computation needed to guarantee that features can be successfully tracked.
In contrast, the MER vehicles 20 MHz CPU and low-throughput camera bus result in
each Visual Odometry update taking a nontrivial amount of time; up to three minutes for
a single tracking step using the nominal mission software. Hence we were driven to the
opposite extreme; take images as infrequently as possible, and make the tracker robust to
large changes in feature location.
Consider also the feature detector. Ensuring that the images used for Visual Odometry
processing contain enough features is still the job of the human rover driver. While planning
a drive that uses Visual odometry, rover drivers need to carefully consider which way to
point the cameras. Much of Spirits driving took place in terrain with so many features that
pointing was rarely an issue. In contrast, year two saw the introduction of the Slip Check
safety constraint on Opportunity to ensure that it would not get bogged down in piles of
sand. The feature detector was optimized for feature-laden natural terrain, and often failed
to find sufficient unique features in the piles of sand, so another approach had to be taken
(happily, the featureless terrain was also very pliable, so the tracks left by the vehicle were
not only visible, but also provided enough features to enable convergence).

In the False Positives discussion in Section 5.4 below, we note that Visual Odometry did
sometimes produce incorrect updates. So we added additional sanity checks during the
second year of operations, to allow the rover drivers to constrain the set of acceptable up-
dates, according to their expectations of the terrain ahead. These optional constraints are
summarized in Table 1.

Robustness to nonconvergence is another difficult issue. A system that updates frequently

can better absorb the impact of a few frames that fail to track. But MER rover drivers need
to write command sequences that closely monitor whether the Visual Odometry processing
has succeeded at each step.

5 Using Visual Odometry on Mars

Visual Odometry processing was performed on both MER rovers using mast-mounted NAV-
CAM imagery. NAVCAMs have a 45-degree field of view and sit 1.5 meters above the ground
plane (Maki et al., 2003), so all Visual Odometry drives were split into steps small enough
to ensure at least 60% overlap between adjacent images. During each step the rover was
typically commanded to drive no more than 75 cm in a straight line or curved arc, and when
turning in place was commanded to change heading by no more than 18 degrees per step.
Motions outside these bounds forced the process to forego any update for that step.

Although Visual Odometry processing could have been beneficial during all rover motion,
each step required an average of nearly three minutes of computation time on MERs 20 MHz
RAD6000 CPU, and thus it was only commanded during specific types of motion: relatively
short drives (typically less than 15 meters) that occurred either on steep slopes (typically
more than 10 degrees), in situations where a wheel was being dragged (digging a trench,
or conserving drive motor lifetime on Spirits right front wheel), or driving through piles of
sand. The onboard IMU exhibited a very small drift rate (usually less than 3 degrees per
hour of operation) and therefore maintained attitude knowledge very well: therefore during
the first two years of operations from January 2004 through January 2006, Visual Odometry
was typically used to update rover position only.

There were some instances in which Visual Odometry did not converge to a solution. These
are mostly attributable to either too large a motion (e.g. commanding a 40 degree turn in
place, resulting in too little image overlap) or to a lack of features in the imaged terrain; but
see the description of False Positives in Section 5.4 too. It has successfully measured slips
as high as 125% on Sol 206 when Spirit tried to drive up a more than 25 degree slope.
Several benefits were realized from Visual Odometry. Vehicle safety was maintained by
having the rover terminate a planned drive early, if it realized via Visual Odometry that
it was making insufficient progress toward its goal vis a Slip Check, or was nearing the
prespecified keep-out location of an obstacle. The improved drive accuracy in new or
mixed-soil terrains also yielded a greater number of science observations, by reducing the
number of sols needed to make targets reachable by the instrument arm (IDD). PANCAM
(Panoramic Camera) and MiniTES science observations requiring precision pointing at a
particular target were often scheduled in the middle of a drive using Visual Odometry, which
eliminated the need for human confirmation of the pointing angle and therefore saved an
extra sol.

5.1 Meridiani Planum: Opportunity Rover

The terrain at Meridiani Planum is a challenging one for Visual Odometry. It is often difficult
or impossible to find a patch of nearby terrain that has enough texture for Visual Odometry
processing to successfully find and track features, because much terrain is covered by a
thick layer of extremely fine particles. Fortunately, areas that have this smooth, featureless
appearance tend to be very flat, and in those areas the IMU and encoder-based position
estimation has performed well enough that Visual Odometry was not needed. Terrain that
exhibits higher slope (and consequently more position uncertainty) almost always has a
distinctive appearance (e.g., bedrock outcrop), or is near enough to interesting features that
Visual Odometry can be employed successfully.

Figure 4: Views of Opportunitys 19 meter drive from Sol 188 through Sol 191. The inside
path shows the correct, Visual Odometry updated location. The outside path shows how
its path would have been estimated from the IMU and wheel encoders alone. Each cell
represents one square meter.

The path predicted by wheel odometry alone can be quite different from the actual path.
Figure 4 shows two views of the trajectory taken by Opportunity during Sols 188-191. The
rover was driven uphill and across slope over a real distance of 19 meters, but wheel odometry
alone would have underestimated it by 1.6 meters and failed to measure the slip-induced
elevation change. The outside path indicates the course as it would have been estimated
solely by wheel odometry (ignoring the translation and additional distance that resulted
from sideslip), and the inside path shows the Visual Odometry-corrected course plot that
was generated onboard and more accurately models actual motion. The final positions differ
by nearly 5 meters.

The earliest benefit from Visual Odometry came inside 20 meter-diameter Eagle Crater,
Figure 5: Wopmay, an obstacle in Endurance Crater 60 cm tall, 90 cm wide, and 150 cm
long.

Figure 6: Looking back at Opportunitys tracks near Wopmay on sol 268 after weeks of
trying to drive away. The slope of the sandy terrain varied from 17 to 23 degrees.

Opportunitys landing site (upper left corner of Figure 15). Most driving inside Eagle Crater
was meticulously planned by human drivers, predicting slip using tables generated by the
mechanical team from Earth-based tests of a rover driving in sand (Lindemann and Voorhees,
2005). But while those tables worked well for predicting purely upslope and cross-slope slips
on pure sand, no model was available for driving on pure bedrock outcrop, mixtures of
bedrock and loose sand, or at angles other than 0, 45 and 90 degrees from the gradient. In
those circumstances Visual Odometry was sometimes used to drive to the proper target, or
ensure that high resolution PANCAM images of science targets taken after a drive would be
pointed right on target.

The most extensive use of Visual Odometry was made by Opportunity inside 130 meter
diameter Endurance Crater from Sol 133 to Sol 312 (see the upper right corner of Figure 15).
Except for a 12 meter approach and return at the lowest point (with lowest rover tilt) on
Sols 201 and 203 and a 17 meter drive on Sol 249, Visual Odometry was used virtually
continuously throughout. Had it not been available onboard, many more sols would have
been needed to approach targets, and fewer targets would have been achieved.

Visual Odometry not only improved target approach efficiency, it also proved crucial to
maintaining vehicle safety. From Sols 249 to 265 Opportunity kept finding itself near a 1.5
meter long rock called Wopmay (see Figures 5 and 7). Although Wopmay was originally
considered a science target, it also proved to be a most difficult obstacle to avoid. It was
Figure 7: Opportunitys 15 sol trajectory near Wopmay (approximately indicated by the
grey ellipse), first driving toward and then trying to get around or away from it. Downslope
is up to the right. In the left plot, the jumps that point up to the right are the result of
Visual Odometry adjusting the vehicles position downslope. Visual Odometry only corrects
the rovers position at the end of each step of less than 1 meter. The right plot shows the
same course with the Visual Odometry jumps removed.

Figure 8: Opportunitys planned 8.7 meter drive along a 20-24 degree slope on Sol 304.

located downhill from a 17-20 degree downslope area comprised of loose sand and buried
rocks. Several attempts to drive around it were thwarted not only by very high slip, but
also by the unseen rocks buried just beneath the surface (one of which caused a stall on sol
265). Fortunately, the human-commanded sequences took into account the possibility that
the rover might slip, and so Opportunity halted several of its planned drives prematurely
(and correctly) when it realized that it was moving too close to Wopmay. The churned up
tracks that resulted from multiple attempts to leave Wopmay can be seen in Figure 6.

Visual Odometry also enabled more precise approaches to difficult targets. On Sol 304, a
drive of over 8 meters was planned on an outcrop whose slope varied from 20 to 24 degrees.
Because the drive plan took a wide range of potential slips into account, Opportunity was
able to drive just far enough across slope, then turn and drive just far enough upslope, to
perfectly position the desired target within the IDD work volume in a single sol. Figure 8
illustrates the planned drive, and Figure 9 shows the final image from the body-mounted
front Hazard cameras (HAZCAMs) showing the target area perfectly located between and
just ahead of the front wheels.

After four months of long distance drives (including 15 sols in which 100 224 meters were
covered in a single sol (Biesiadecki and Maimone, 2006)), during which Visual Odometry
Figure 9: After Opportunitys 8.7 meter slope drive on Sol 304, the goal area is perfectly
reachable inside the IDD work volume, indicated in the green shaded area in the middle of
the image.

Figure 10: Our first close-up first view of Purgatory ripple from sol 446. The churned up
tracks near the rover are only 2 meters long, but the wheels had rotated enough to have
traveled 50 meters on open terrain. This event led us to use the Slip Check drive mode for
all future drives across sandy terrain.

was rarely used, it suddenly came to the fore again. On sol 446, Opportunity nearly buried
its wheels in Purgatory ripple, a nondescript pile of sand, by executing 50 meters of blind
driving while failing to climb over it (see Figure 10). How could we get out of it, and how
could we stop ourselves from doing the same thing in the next pile of sand when we did
finally break free? After much deliberation, a solution was proposed that hinged on Visual
Odometry. After each commanded 2 meters of driving, Visual Odometry would be used
to estimate the vehicles motion. If Visual Odometry confirmed that little or no motion
had taken place, more commands would be sent. But once Visual Odometry measured a
nontrivial motion or failed to converge, all drive commands would stop so we could take a
close look back at the pile that had caused so much trouble (see Figure 11).

Experiments verified that although the Visual Odometry feature detector often failed to find
enough features when looking at piles of sand, there was a set of features rich enough to enable
Figure 11: Looking back on sol 491 at Opportunitys tracks on and near Purgatory ripple,
after having spent a month escaping it.

Table 2: Assessment of 256x256 NAVCAM-based Onboard Visual Odometry during Purga-

tory extraction. For each sol, we considered three points in 3D: start position, end position
as estimated onboard, and end position as estimated on Earth (Ground) using Hi-Resolution
PANCAM images. Columns indicate activity Sol, number of NAVCAM Visual Odometry
updates computed onboard, start to Onboard end position magnitude, start to Ground end
position magnitude, vector between end positions magnitude, vector between end positions
magnitude divided by number of updates.
Num. of Onboard Ground Hi-res Average
Sol Onboard Distance Distance Pancam Error per
Updates Estimate Estimate Correction Step
463 10 0.029 m 0.030 m 0.0084 m 0.0008 m
465 10 0.020 m 0.020 m 0.0090 m 0.0009 m
466 10 0.033 m 0.026 m 0.0121 m 0.0012 m
467 10 0.029 m 0.021 m 0.0196 m 0.0020 m
468 20 0.037 m 0.039 m 0.0125 m 0.0006 m
469 5 0.010 m 0.011 m 0.0026 m 0.0005 m
470 6 0.059 m 0.057 m 0.0053 m 0.0009 m
471 6 0.063 m 0.061 m 0.0071 m 0.0012 m

successful estimation: the rovers own tracks. This strategy proved highly successful. From
sol 463 through sol 483, onboard Visual Odometry measured progress in the millimeters; slip
rates varied from 98.9% to 99.5%. These numbers were confirmed both by manual inspection
of the images, and also by running Visual Odometry processing on Earth using images from
Opportunitys high-resolution PANCAM stereo pair taken before and after each sols drive;
progress on the order of 1 millimeter of actual motion vs 2 meters commanded (with wheels
steered to perform a 23 degree heading change) was seen for many sols. Finally on sol 484,
all driving commands were terminated when Visual Odometry failed to converge three times
indicating that it had driven so far that features could not be tracked.

From this point on, a new driving strategy was adopted for Opportunity. Blind drives over
sandy terrain would be limited to some maximum distance, initially 5 meters. After each
5 meter drive segment, Visual Odometry would be used to perform a Slip Check (running
Visual Odometry over a short, 20 cm drive step) to ensure that the rover was still capable
of forward motion. If the amount of forward progress was less than some minimum fraction,
all driving would be stopped. The Slip Check has stopped several drives from digging in
(e.g., sols 501, 603), and thus enabled driving to resume one sol later instead of the 39 sols
required to exit Purgatory.

Visual Odometry results are summarized in Table 3. As of February 2006, Opportunity

has thrived for 720 sols. Visual Odometry was used more here than on Spirit, because
Opportunity spent more of its first year on slippery surfaces. It has converged to a solution
95% (828/875) of the time.

5.2 Performance Assessment

We will not attempt to fully characterize the accuracy of the updates computed by Visual
Odometry over multiple long drives on Mars, because our means of measuring what actually
happened is extremely limited. Publications such as those cited in Section 3 provide the best
performance measurements for the algorithm, albeit in Earth-based testbeds and without
all the additional constraint checks. However, the extraction from Purgatory does provide a
limited means of evaluating the performance of onboard Visual Odometry.

In general we have no ground truth against which to compare the position estimates calcu-
lated from the 256x256 NAVCAMs. But during the drives that were performed to escape
from the Purgatory ripple, not only were NAVCAM images of the tracks 1.3 4 meters away
taken between each drive step, high-resolution 1024x1024 PANCAMs of a single track 2.5
4 meters away were also taken at the beginning and end of each sols multi-step drive. The
actual daily distance driven during the first eight sols was never more than 7 cm, so while
these drives are not representative of the mission as a whole, they do give us a means of esti-
mating the average error introduced by running Visual Odometry onboard. By treating the
results from high-resolution PANCAM images as an independent estimate of the distance
driven each sol (ground truth), we can estimate the precision with which the onboard
estimates are computed.1 We performed Visual Odometry processing on Earth using the
PANCAM images, and then compared the delta computed from each sols before-and-after
high-resolution 1024x1024 images against the cumulative onboard estimate computed from
multiple low-resolution 256x256 NAVCAM pairs. The magnitude of the correction com-
puted from the high-resolution PANCAM pair divided by the number of updates generated
onboard gives us an average error associated with each Visual Odometry update that was
run onboard during a particular sol. Table 2 summarizes the data, which demonstrates that
the average error from a single update was less than 2 millimeters during each of the eight
sols in the table, which is within the measurement error of the ground truth.

In a qualitative sense, we know that our algorithm performs well because the rovers be-
havior matches our predictions better when Visual Odometry is enabled. To truly measure
the performance of our algorithm in nominal Martian terrain, we would need a good source
of ground truth and substantial data reduction efforts. Possible sources of this informa-
tion include MER stereo views of the tracks left by Visual Odometry-enabled drives (with
potential precision at the centimeter level, but difficult to correlate with each step), post-
drive bundle-adjusted localization estimates such as those summarized in (Li et al., 2005;
Li et al., 2006) (of various precisions), and orbital images (e.g., 1.5 meter resolution images
1
We estimate the stability of the high-resolution PANCAM updates themselves by comparing PANCAM images
taken while the rover was motionless (comparing each end-of-sol PANCAM against the next sols beginning-of-sol
PANCAM). Except for lighting differences such as shadows, these views should have been identical. Over seven of
the same eight sols, the mean computed drift was 1.3 mm +/- 0.5 mm, with a maximum drift of 2 mm.
Figure 12: Spirit spent sols 339 through 345 dealing with a potato-sized rock that got stuck
in the right rear wheel (viewed here on the left side of this image from the rear hazard
camera).

from Mars Global Surveyor, and 0.25 meter resolution images from Mars Reconnaissance
Orbiter). Good candidate locations for extensive use of Visual Odometry on Opportunity
would be Endurance Crater and the area around Purgatory ripple, and on Spirit the initial
exploration of the Columbia Hills (in general, any dense blue area in Figures 14 and 15).
Additional performance quantification may be achievable from these resources, but that is
left to future work. Raw images from the rovers are available on the Mars Rovers web site.

5.3 Gusev Crater: Spirit Rover

The terrain at Gusev crater is well suited for Visual Odometry processing. The rock abun-
dances there matched predicted distributions (Golombek and Rapp, 1997), resulting in a
generally feature-rich landscape with detailed textures comprised of rocks of different sizes
and brightnesses. When planning for drives using Visual Odometry, rover drivers typically
only had to bear in mind the restriction that adjacent frames should have at least 60% im-
age overlap, only occasionally having to avoid pointing the cameras at infrequently occurring
sand dunes. As a result, Spirits Visual Odometry software has performed well.

One unique driving mode that benefited a great deal from Visual Odometry on Spirit was
wheel-dragging. The right front wheel was found to draw more current while driving than
any of the other wheels starting on Sol 125. This concern led to the development of a driving
strategy to conserve motor lifetime, during which that wheel would be dragged while all the
others were driven. Although this was found to enable reasonable progress on relatively
flat terrain, error in the position estimate grew substantially in this mode. The Visual
Odometry capability meant that not only could progress be made, but also the error added
to the onboard position estimate could be bounded as well.

Another invention by the rover drivers was a Visual Odometry-based Obstacle Check style of
driving (Biesiadecki et al., 2007). Human rover drivers could create a set of keep-out zones
Figure 13: Image of Opportunitys shadow on Sol 235, with tracked features overlaid. Red
dots indicate the features in this current image, blue dots and vectors show the locations of
corresponding features found in the previous image,

around whatever was perceived to be a nearby obstacle. By driving with Visual Odometry
enabled all the time, we could ensure that the rover would stop driving if it ever slipped
too close to any obstacle by checking the rovers corrected position against the preset list
of obstacle locations. This style of driving was used whenever commanded drives took the
rover uphill of any obvious obstacles. Although the rover was capable of detecting rocks and
geometric hazards on its own (Biesiadecki and Maimone, 2006), this human-guided obstacle
detection was used to allow drives to execute more quickly and hence cover larger distances.
Also, the MER vehicles have no means of detecting non-geometric hazards (like slippery
sandy areas) onboard at all. Whenever there was a concern about such non-geometric
hazards, human drivers had to manually create a keep-out zone around them. The mobility
sequence of commands would then poll the rovers location and compare it to the list of
keep-out areas, halting the drive if it got too close.

On sol 339 Spirits right rear wheel stalled when a potato-sized rock that had been buried
under soil got wedged in that wheel (see Figure 12). Up to this point, such small rocks had
posed no hazards to mobility, and Visual Odometry had been used only as a tool for improv-
ing overall position knowledge. But the experience of having to spend a week extricating
the potato-sized rock led the rover drivers to create a new use for Visual Odometry: as a
Slip Checking device. Rather than worry about the overall position, Visual Odometry could
be used infrequently to measure how much relative slip the rover had just encountered; for
example, if a 40cm arc had been commanded but fewer than 20cm of progress was measured,
then it would be clear that the rover was slipping more than 50%. The sol 339 event led to
Slip Checks becoming a standard speed optimization; instead of driving with Visual Odome-
try running all the time (with a max drive rate of only around 10 meters/hour), Slip Checks
could be used to place an upper bound how deeply we might get dug in (e.g., by limiting
blind drive segments to 5 meters).

Relatively little slip was seen during the first six months of Spirits mission. But once the
base of the Columbia Hills was reached, drives up into the hills were found to exhibit much
more unpredictable slip. Thus Visual Odometry has been used during most of the drives
in the Columbia Hills, especially to ensure that Spirit stays far enough away from nearby
rock obstacles. The average tilt of the rover during those times that Visual Odometry was
commanded was 14.4 degrees +/- 4.4 degrees, counting 625 samples spanning an absolute
range from 2 - 30 degrees.
Visual Odometry results are summarized in Table 3. As of March 2005, Spirit has thrived
for 414 sols. Visual Odometry was only used on Spirit after it had reached the Columbia
Hills, nearly six months into its mission. But since then it has converged to a solution 97%
(590/609) of the time.

Table 3: Results of running Visual Odometry (Visodom) on Mars. Expressions ms indicate

mean m and standard deviation s.
Spirit Opportunity
Lifetime as of 4 March 2005 414 sols 394 sols
Total Drive Distance 4161 meters 3158 meters
Days Spent Driving 184 sols 172 sols
Days Using Visual Odometry 52 sols 75 sols
Nominal Evaluation Steps 609 pairs 875 pairs
Nominal Initialization Steps 57 pairs 75 pairs
Forced Init by Large Turn 10 pairs 11 pairs
Forced Init by Planned Repointing 5 pairs 9 pairs
Forced Init by Driving Too Far 1 pairs 2 pairs
Total Visodom Image Pairs 682 pairs 972 pairs
Successful (non-initial) Convergences 590 pairs 828 pairs
Iterations (assume Convergence) 6.4 1.7 iterations 8.4 5.2 iterations
Features Tracked at each Step 73.4 29.3 features 87.4 34.1 features
Non-convergences 19 pairs 47 pairs
Mean Updates per Drive Sol 12.0 8.5 pairs 12.9 10.2 pairs
Max Updates per Drive Sol 33 pairs 59 pairs
Mean Rover Tilt During Visodom 14.6 4.4 degrees 18.0 4.6 degrees
Rover Tilt Range During Visodom 2 - 30 degrees 0.8 - 31 degrees

5.4 False Positives

Although we had never seen Visual Odometry converge to an inaccurate solution during
testing, on Opportunity Sols 137 and 141 several unreasonable position updates were com-
puted onboard. These are attributable to an improper parameter setting; at that time, the
minimum separation between features was too small. As a result, the set of detected features
was allowed to cluster tightly around a small planar but feature-rich area. Increasing that
parameter was all that was needed to allow the software to find additional features with
better spatial distribution and then converge to a reasonable solution.

The only other instance of a false positive solution in this part of the mission was on Sol
235. During that sol the NAVCAMs were pointed at two small and widely separated rocks.
Although features were found on those rocks, more usable features were found on the largest
shape in the image; the rovers shadow (see Figure 13). Even though forward drive progress
was made, the onboard estimator assumed that the shadow (having more consistently tracked
features spread throughout the image) better reflected actual motion, and therefore produced
an incorrect estimate. This problem would not have arisen had there been more interesting
texture around and under the shadow, so since that sol human drivers have had to take the
rover shadow into account whenever planning Visual Odometry drives.
5.5 Suggested Improvements

The MER mission has demonstrated the usefulness of Visual Odometry technology for plan-
etary exploration. But several open issues remain to be resolved for future missions.

Runtime The MER system architecture was designed for robustness, not speed. Image ac-
quisition alone could take tens of seconds, including the 2.5 second acquisition of a 256x1024
binned pair of images (multiplied by however many images are needed for auto-exposure
processing), plus tens of seconds more to prepare the image for potential transmission as
a separate data product to Earth. The Visual Odometry processing itself took place on a
single 20 MHz CPU that shared system resources between the intensive image processing
operations and dozens of realtime tasks. One complete processing cycle from image acqui-
sition through the position update took 3 minutes on average, limiting overall drive speed
to approximately 10 meters/hour (this is expected to improve somewhat in the September
2006 flight software update). Faster processing would allow more frequent use of Visual
Odometry, improving vehicle safety, predictability, and autonomous capability.

Finding Good Terrain The implicit assumption in any Visual Odometry algorithm is that
the terrain exhibits a rich visual texture, but MER encountered many areas with little or no
easily-tracked texture, on both rovers. Spirits terrain at Gusev crater typically had an in-
teresting appearance that made tracking easy, but occasional sandy areas caused problems.
Opportunitys terrain at Meridiani was the opposite, in that most of it lacked trackable
features. To become a truly autonomous capability, Visual Odometry would need to be ex-
tended to autonomously (and efficiently) search for appropriately textured terrain, instead of
requiring human operators to select the viewing direction. The search space might include
physical changes like re-pointing the cameras or switching lens filters, or algorithmic im-
provements to detect or eliminate shadows, dynamically update the imaging parameters to
adapt to the terrain, and switch from feature-based to area-based homography or shape reg-
istration methods autonomously (e.g., (Olson and Matthies, 1998)) (NAVCAM-based dense
stereo processing works well even on the sandy terrain that exhibits no obvious features).

Wide-angle lenses Our feature tracker was not able to track features that underwent large
scale changes between images. This prevented us from using wide-angle HAZCAM images
for position estimation on Mars, even though they seemed to work well on Earth; although
HAZCAM tests worked in our indoor test area, we later determined that the indoor success
was largely due to having tracked features on the walls, which did not change scale between
motions. With no walls on Mars, and a desire to maximize the distance between images to
enable fast driving, the HAZCAMs were eliminated as a useful sensor for Visual Odometry
on MER. A future ability to use the wide-angle HAZCAMs could save time if HAZCAMs had
already been acquired for autonomous obstacle assessment, and would also make it possible
to see more of the terrain (and hence find more trackable features without having to repoint
cameras). Lowes work on scale invariant features is one possible approach (Se et al., 2002).

Flexibility Maintaining program flexibility is critical in adapting technology to novel domains,

and Mars is full of surprises. For instance, although originally tested on 128x128 HAZCAM
images, Visual Odometry was found to work best using 256x256 NAVCAMs. Also, the rover
often had to be driven backward to give the NAVCAMs an unobstructed view of the tracks,
since rover tracks provided the most viable features in the sands of Meridiani Planum. Most
spaceflight control software is written with an eye toward reducing complexity, but flexibility
is paramount when new terrains and new surprises are found every sol.
Figure 14: Plot of Spirits traverse history using Visual Odometry in the Columbia Hills from
sols 1 850. Units are in meters from the landing site origin, as measured onboard the rover.
Cyan lines indicate directly commanded blind drives, red lines indicate blind drives with
autonomous heading compensation, green lines indicate autonomous hazard detection, and
blue lines indicate Visual Odometry. Spirit only used Visual Odometry within the Columbia
Hills, not during its three kilometer trek to reach them.

6 Conclusion

Visual Odometry has been a highly effective tool for maintaining vehicle safety while driving
near obstacles on slopes, achieving difficult drive approaches in fewer sols, performing Slip
Checks to ensure the vehicle is still making progress, and ensuring accurate science imaging.
Although it requires active pointing by human drivers in feature-poor terrain, the improved
position knowledge enables more autonomous capability and better science return during
planetary operations.

For planetary exploration applications, the primary performance limitation for Visual Odom-
etry at present is the runtime of the image processing part of the task given the highly
constrained flight computers available, not accuracy or precision of the motion estimates.
Therefore, addressing this runtime bottleneck is the overriding immediate priority for future
work in this domain, through any combination of algorithmic improvements and advances in
flight computers. In the longer term, more robust feature tracking and autonomous terrain
selection would be needed to make it a truly robust technology.

7 Acknowledgments

Thanks to Jeff Biesiadecki for integrating this capability into the MER flight software Mobil-
ity Manager state machine, the Surface Navigation and Mobility testing team for testing and
Figure 15: Plot of Opportunitys traverse history using Visual Odometry from sols 1 830.
Opportunity landed in 20 meter diameter Eagle crater in the upper left, drove in and around
Endurance crater (upper right) from sols 133 312, and continued to drive south toward
Victoria crater as of sol 830. Units are in meters from the landing site origin, as measured
onboard the rover. Cyan lines indicate directly commanded blind drives, red lines indicate
blind drives with autonomous heading compensation, green lines indicate autonomous hazard
detection, and blue lines indicate Visual Odometry.

validating it in time to be included the mission, flight software lead Glenn Reeves for sup-
porting its inclusion, Steve Goldberg, Dan Helmick, Clark Olson and Marcel Schoppers for
their contributions to algorithm development and implementation, the RSVP team (Brian
Cooper, Frank Hartman, Scott Maxwell, John Wright and Jeng Yen) for the 3D rover course
visualizations, and Robert Liebersbach for the course plot and corrected pose plotting tools.

The work described in this paper was carried out at the Jet Propulsion Laboratory, Cal-
ifornia Institute of Technology, under a contract to the National Aeronautics and Space
Administration.

References

Amidi, O., Kanade, T., and Fujita, K. (1999). A visual odometer for autonomous
helicopter flight. Journal of Robotics and Autonomous Systems, 28:185 193.
http://www.ri.cmu.edu/pub files/pub4/amidi omead 1999 1/amidi omead 1999
1.pdf.
Biesiadecki, J. J., Baumgartner, E. T., Bonitz, R. G., Cooper, B. K., Hartman, F. R.,
Leger, P. C., Maimone, M. W., Maxwell, S. A., Trebi-Ollenu, A., Tunstel, E. W., and
Wright, J. R. (2005). Mars Exploration Rover surface operations: Driving Opportunity
at Meridiani Planum. In IEEE Conference on Systems, Man and Cybernetics, The Big
Island, Hawaii, USA.
Biesiadecki, J. J., Leger, P. C., and Maimone, M. W. (2007). Tradeoffs between directed and
autonomous driving on the Mars Exploration Rovers. International Journal of Robotics
Research. Special Issue on ISRR 2005, to appear.
Biesiadecki, J. J. and Maimone, M. W. (2006). The Mars Exploration Rover surface mobility
flight software: Driving ambition. In IEEE Aerospace Conference, volume 5, Big Sky,
Montana, USA.
Corke, P. I., Strelow, D., and Singh, S. (2004). Omnidirectional visual odometry for a plane-
tary rover. In Proceedings of IROS 2004. http://www.ri.cmu.edu/pubs/pub 4913.html.
Gluckman, J. and Nayar, S. K. (1998). Ego-motion and omnidirectional cameras. In ICCV,
pages 9991005.
Golombek, M. and Rapp, D. (1997). Size-frequency distributions of rocks on Mars and Earth
analog sites: Implications for future landed missions. Journal of Geophysical Research -
Planets, 102(E2):41174129.
Helmick, D., Cheng, Y., Roumeliotis, S., Clouse, D., and Matthies, L. (2004). Path following
using visual odometry for a Mars rover in high-slip environments. In IEEE Aerospace
Conference, Big Sky, Montana, USA.
Lacroix, S., Mallet, A., Chatila, R., and Gallo, L. (1999). Rover self localization in
planetary-like environments. In International Symposium on Artificial Intelligence,
Robotics, and Automation for Space (i-SAIRAS), pages 433440, Noordwijk, Nether-
lands. ftp://ftp.laas.fr/pub/ria/simon/isairas99.ps.gz.
Li, R., Archinal, B. A., Arvidson, R. E., Bell, J., Christensen, P., Crumpler, L., Marais,
D. J. D., Di, K., Duxbury, T., Golombek, M., Grant, J., Greeley, R., Guinn, J.,
Johnson, A., Kirk, R. L., Maimone, M., Matthies, L. H., Malin, M., Parker, T.,
Sims, M., Thompson, S., Squyres, S. W., and Soderblom, L. A. (2005). Spirit rover
localization and topographic mapping at the landing site of Gusev Crater, Mars.
JGR-Planets, Special Issue on Spirit Rover, 111(E02S06, doi:10.1029/2005JE002483).
http://www.agu.org/journals/ss/SPIRIT1/.
Li, R., Arvidson, R. E., Di, K., Golombek, M., Guinn, J., Johnson, A., Maimone, M.,
Matthies, L. H., Malin, M., Parker, T., and Squyres, S. W. (Submitted 2006). Opportu-
 nity rover localization and topographic mapping at the landing site of Meridiani Planum,
Mars. Journal of Geophysical Research - Planets.
Lindemann, R. A. and Voorhees, C. J. (2005). Mars Exploration Rover mobility assembly
design, test and performance. In IEEE Conference on Systems, Man and Cybernetics,
The Big Island, Hawaii, USA.
Maki, J. N., III, J. F. B., Herkenhoff, K. E., Squyres, S. W., Kiely, A., Klimesh, M.,
Schwochert, M., Litwin, T., Willson, R., Johnson, A., Maimone, M., Baumgartner,
E., Collins, A., Wadsworth, M., Elliot, S. T., Dingizian, A., Brown, D., Hagerott,
E. C., Scherr, L., Deen, R., Alexander, D., and Lorre, J. (2003). Mars Exploration
Rover engineering cameras. Journal of Geophysical Research, 108(E12):1211224.
http://www.agu.org/pubs/crossref/2003/2003JE002077.shtml.
Matthies, L. (1989). Dynamic Stereo Vision. PhD thesis, Carnegie Mellon University Com-
puter Science Department. CMU-CS-89-195.
Matthies, L. and Shafer, S. (1987). Error modelling in stereo navigation. IEEE Journal of
Robotics and Automation, RA-3(3).
McCarthy, C. and Barnes, N. (2004). Performance of optical flow techniques for indoor nav-
igation with a mobile robot. In International Conference on Robotics and Automation,
pages 50935098, New Orleans, USA.
Moravec, H. (1980). Obstacle Avoidance and Navigation in the Real World by a Seeing Robot
Rover. PhD thesis, Stanford University, Stanford, CA.
Nister, D. (2004). An efficient solution to the five-point relative pose problem. IEEE Trans.
Pattern Anal. Mach. Intell., 26(6):756777.
Nister, D. (2005). Preemptive ransac for live structure and motion estimation. Machine
Vision and Applications, 16(5):321329.
Nister, D., Naroditsky, O., and Bergen, J. (2004). Visual odometry. In IEEE Conference
on Computer Vision and Pattern Recognition, pages 652 659. IEEE Computer Society
Press.
Nister, D., Naroditsky, O., and Bergen, J. (2006). Visual odometry for ground vehicle
applications. Journal of Field Robotics, 23(1):320.
Olson, C. F. and Matthies, L. H. (1998). Maximum likelihood rover localization by matching
range maps. In International Conference on Robotics and Automation, pages 272277.
Olson, C. F., Matthies, L. H., Schoppers, M., and Maimone, M. W. (2003). Rover navigation
using stereo ego-motion. Robotics and Autonomous Systems, 43(4):215229.
Schonemann, P. H. and Carroll, R. M. (1970). Fitting one matrix to another under choice
of a central dilation and a rigid motion. Psychometrika, pages 245255.
Se, S., Lowe, D., and Little, J. (2002). Mobile robot localization and mapping with uncer-
tainty using scale-invariant visual landmarks. International Journal of Robotics Research,
21(8):735758.
Vassallo, R. F., Santos-Victor, J., and Schneebeli, J. (2002). A general approach for egomo-
tion estimation with omnidirectional images. In OMNIVIS, pages 97103.
Zhang, Z., Faugeras, O., and Ayache, N. (1988). Analysis of a sequence of stereo scenes con-
taining multiple moving objects using rigidity constraints. In International Conference
on Computer Vision, pages 177186, Tampa, Florida. Computer Society of the IEEE,
IEEE Computer Society Press.