IMU and Cable Encoder Data Fusion for In-Pipe

Mobile Robot Localization

Andreu Corominas Murtra, Josep M. Mirats Tur

CETaqua, Water Technological Center. Barcelona, Catalunya, Spain
www.cetaqua.com.acorominas/[email protected]

Abstract-Inner pipe inspection of sewer networks is a hard This paper proposes and evaluates a complementary ap­
and tedious task, due to the nature of the environment, which proach to localize a mobile robot, which is based on data
is narrow, dark, wet and dirty. So, mobile robots can play an fusion of two sensor devices: an IMU and a cable encoder,
important role to solve condition assessment of such huge civil the later measuring the length of an unfolded cable, from the
infrastructures, resulting in a clear benefit for citizens. One of the
starting point of operations up to the robot platform, which is
fundamental tasks that a mobile robot should solve is localization,
but in such environments GPS signal is completely denied, so
tethered for safety, communications and energy reasons. This
alternative methods have to be developed. Visual odometry and
sensor set-up has the great advantage of its availability, since it
visual SLAM are promising techniques to be applied in such is independent of environment situations which usually cause
environments, but they require a populated set of visual feature failures in image processing. The presented method is based
tracks, which is a requirement that can not be fulfilled in such on the Graph SLAM approach [13], [1], [4], so it can be easily
environments in a continuous way. With the aim of designing exported as a complementary method to add robustness to an
robust and reliable robot systems, this paper proposes and inertial/vision based Graph SLAM estimation. Optimization is
evaluates a complementary approach to localize a mobile robot, carried out within an interval of consecutive platform states,
which is based on sensor data fusion of an inertial measurement called a window state, given the sensor measurements in that
unit and of a cable encoder, which measures the length of an
interval. Optimization tries to find out the most likely window
unfolded cable, from the starting point of operations up to the
state, given sensor measurements, where likelihood is treated
tethered robot. Data fusion is based on optimization of a set of
windowed states given the sensor measurements in that window.
in a probabilistic way. In terms of a cost function, such opti­
The paper details theoretical basis, practical implementation mization searches the window state point that minimizes a cost
issues and results obtained in testing pipe scenarios. function, which is expressed as a sum of several constraints,
imposed by measurements. The paper details theoretical basis,
practical implementation issues and results obtained in testing
I. INTRODUCTION
pipe scenarios with a wheeled platform.
Inner pipe inspection of sewer networks is a hard and
tedious task, due to the nature of the environment, which II. STATE DEFINITIONS AND GRAPH MODEL
is narrow, dark, wet and dirty. Mobile robots can play an We consider the following full platform state at iteration t:
important role to solve condition assessment of such huge
civil infrastructures, resulting in a clear benefit for citizens. (1)
For instance, only at the city of Barcelona, sewer network has
an overall length of 5000 K m, an underground asset that has which is actually a column vector where pt [p� pt p�l T
grown over the last centuries, built of different materials and
sizes. Inspection robots have a big market to be exploited in
is the position of the robot, vt = f
[v� vt v; is the velocity
and qt [q6 qi q� q�lT is the quaternion representation of the
=

such scenarios, but fundamental robot capabilities have to be

vehicle orientation, being q6 the real part of the quaternion. All
robustly solved to definitively introduce them as reliable and
three vectors are expressed with respect to a given reference
market-ready solutions.
frame, which can be the starting point of operations.
One of these fundamental tasks that a mobile robot should Position part, pt, which is the final goal variable to be
solve is localization, taking into account that in such envi­ estimated, will be computed by integrating the optimized linear
ronments GPS signal is completely denied. To overcome this velocities, out of the minimization loop, since neither IMU nor
limitation, alternative methods have to be developed. Visual cable encoder measure the position component of the state
odometry [12], [9], [5] and visual SLAM [2], [7], [11] are xt (they measure derivatives or indirect values). Therefore we
promising techniques to be applied in such environments. redefine the platform state in a dimensional-reduced way, much
Adding an Inertial Measurement Unit (lMU) to the sensor more efficient to be used by an optimization procedure as it
set-up provides extra robustness, while it overcomes scale ob­ will be explained on section III. The so called platform state
servability lack of monocular vision approaches [1], [10], [8]. is:
However, such vision based methods require a populated set of xt [vtx vty vtz qt1 q2t qt3 1 T .
= (2)
visual feature tracks to robustly estimate robot positions. This
requirement is not allways fulfilled in inner pipe scenarios, due Only velocity components and imaginary part of the quaternion
to areas of poor pipe surface texture, to sudden water flows have been kept, since they are the independent values to be
that difficults image processing steps, or to high illumination estimated directly. Real part of the quaternion can be computed
reflections produced by on-board light source. by forcing Iqtl =1.

978-1-4673-6225-2/13/$31.00
Authorized ©2013 IEEE
licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on September 08,2023 at 18:07:29 UTC from IEEE Xplore. Restrictions apply.
 The trajectory run by the robot is modelled as a graph where the first term FJ ( si ) is the cost function associated to
linking consecutive nodes, where each node represents a plat­ IMU measurements, the second one FC ( si ) is that associated
form state at a given iteration t, and each link sets a constraint to cable encoder data and the third one, FW ( si ) is that
imposed by sensor data between two of such nodes. Figure 1 imposing the wheeled vehicle constraints. These three function
shows the raw graph issued from this model, composed by costs are detailed in the following section III. Please note that
N + 1 platform states. F ( si ) does not have an explicit expression, but given a set of
sensor measurements, and given a point si, we can compute
Since IMU data is provided at much higher rate than cable its value (pointwise evaluation), so numerical techniques will
encoder measurements, localization will be computed for a be employed to compute the required minimization. The goal
subset of those platform states depicted on figure 1, so a for each optimization is to find the s� point, which minimizes
new version of the graph, called reduced graph, is thought the function F ( si ) :
which only incorporates such selected platform states, also
called keyframes. Figure 2 shows the reduced graph, where (5)
all nodes are linked by a cable encoder measurement and by
an integration of a subset of consecutive IMU measurements. A. IMU Constraint
Based on this reduced graph we will formally define a window
as a subset of Nw + 1 consecutive platform states (i.e. Nw + 1 An IMU device provides high rate (up to 1 KHz) inertial
consecutive keyframes). measurements. At iteration t, the IMU measurement is:
T
So at this point three iteration index can be formally "t =
[0/ 0/ 0/ wt wt wt]
x y z X Y Z '
(6)
defined. Assuming that IMU is the highest rate sensor, the first T
one is the index running on IMU iterations, t E [0, N]. The where at [ a� at a ;] is the body acceleration vector
=

second index is that of those iterations marking keyframes, T

and wt [w� wt w;] is the rotational rate vector, both
=

i E [0, NK], and the third index runs over a window state, measurements reported with respect to the current vehicle
j E [0, Nw]. A window state is defined as: coordinate frame. With these measurements, a set of con­
si =
[xi xH1 ... xHj ... xHNw] , (3)
straints can be imposed to the pose graph, but these constraints
are not formulated on such measurement space. Alternatively,
which is also a column vector concatenating Nw + 1 platform consecutive measurements are integrated to compute the so
states, from xi to xHNw. called integrated platform states, labelled as xJ. Therefore,
for the ith window, the following vector is defined:

(
Please note that t, i and j indexes are integer values. Each
of them has a corresponding time stamp value labelled, respec­
tively, as Tt ,T and THj, allowing proper ordering operations
o� = [ ( x}+1 _ xH1 )T . . . ( x}+Nw _xHNw ) T (7)
i
between them. where terms xHj,Vj E [1, Nw] are provided by the evaluation
point si, while x}+j,Vj E [1, Nw] are computed with xi initial
III. STATE OPTIMIZATION (anchored) point and IMU measurements {"tITt E [T ' THNw]}'
i
The cost function to be minimized associated to IMU measure-
State estimation will be computed for each window over the
ments is:
window state space defined by si. However, we assume the first
state of each window, xi, as a starting point, so it is anchored (8)
and optimization will not modify it. The cost function to be so at this point we have to provide a computer routine that,
minimized is: given measurements during the window period, and given
an evaluation point si, is able to output FJ ( si ) . Therefore,
(4)
we have to detail computation of x}+j, and also matrix C},
which parameterizes the uncertainty of IMU measurements,
thus (C})-1 weights the overall IMU constraint.
Within the ith window, the first integrated state, x} , is
provided by the first (anchored) platform state of the evaluation
point si, and the following integrated platform states, x} are
Fig. 1. Raw graph created when fusing IMU and cable constraints. computed with the following equations:
Black arrows are IMU measurements and red ones represent cable length
measurements from the starting motion point. v} vj-1 + Rt-1(at bQJ�t gO�t ,
= _ _

qtJ- WtqtJ-1 '

_

(9)
q} +-
qt
q
I I
�
(quaternion normalization) ,

where �t is the time elapsed between time stamps Tt -1 and

Fig. 2. Reduced graph with NK + 1 keyframe platform states. Each pair of Tt, Rt-1 R(qj-1) is the rotation matrix associated to the
=

neighboring nodes are linked by two constraints: (I) Black, IMU constraints
imposed by integrating several consecutive inertial measurements, and (2) Red,
platform orientation at t - 1, gO is the gravity vector, and
cable length constraints imposed as the difference of two consecutive cable b" and bw are the accelerometer and gyro bias respectively.
encoder measurements. Gravity, and bias are estimated initially as it is explained in

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on September 08,2023 at 18:07:29 UTC from IEEE Xplore. Restrictions apply.
 subsection III-D. Recall that from quaternion representation, Algorithm 1 IMU Cost Function

]
rotation matrix is: INPUTS
·WINDOW STATE POINT: si
qOq2 +qlq3 ·SENSOR DATA: {{t ,\fth E h,THNw]}
q2q3 - Qq l ,
1 2 O 2 ·KEYFRAME TIMESTAMPS: {THj,\fj 1 . . . Nw} =

2 Q1 Q2
_ _

·PARAMETERS: g,O ba,bw,aa ,wa

(10)
OUTPUT: FI(si)
and matrix Wt is the small angle approximation of the
measured rotational increment wtD.t (typically D.t < 10-2 s): xf = xi Ilanchored platform state, xi = si [1 : 6]
Cf = [06]
-(w�-bwro) -(w�-bwy) ;
-(W -bwz) 1 j=1
Fr = 0
2/ iJ.' (w;-bwz) -(wy-bwy)
while Tt < Ti+Nw do
-(w;-bwz) 2/ iJ.' (w�-bw�) . t
xf, Cf f- imulntegration(xf, Cf,L ,gO, ha, hw)
(w�-bwy) -(w�-bw�) 2/ iJ.' t++
(11)
if Tt > T +j then
i
+ +
On the other hand, matrix C} will weight the constraint, Fr+= (xf - xi j)TC.rl(Xf - xi j)

based on the estimated uncertainty, which depends on the j++

model and on the noisy measurements. So matrix C} will be end if
end while
of the form:
return Fr

(12) which is the difference between measured cable lengths at

the end and start points of the window. Given an evaluation
point si, an expected length difference, gi, can be computed as
the sum of the forward vehicle increments within the window,
where being UF the normal forward vector with respect to the vehicle
-J Cft-1JT + J2 QJ,
Cft - 1 1 / 2T (13) frame:
Nw-l
with JacobianJ1: gi L uF(RHj)-lvHj D.�!;+l, (18)
], j= O
=

J1,vq (14)
J1,qq which implicitly assumes that cable deployment follows the
inner pipe geometry, and primary force that pulls the cable
deployment is due to forward vehicle motion. D.�!�+l is the

J2 =
ax�
at}
=
[ J2,0 va3 0 3
J2,qw
]' (15)
ellapsed time between consecutive keyframes i + j and i + j + 1.
Given cable measurements and the evaluation point si, the
difference between expected, gi, and measured, pi, is:
and matrix QI is the IMU device noise matrix:
t5b(si,Ai,AHNw) = gi(si) - gi(Ai,AHNw). (19)
(16)
At this point, we could follow the same approach described
Jacobian matrices J1,vq, J1,qq, J2,va and J2,qw are partial by IMU function cost, but we realise that, within window
derivatives of v} and q} with respect to q�-I, at and wt. intervals, noise of cable measurement differences was mainly
due to resolution error, since encoder measures up to lern
Algorithm 1 summarizes the procedure to compute the cable length variations. This fact inspires us an alternative
IMU part of the cost function. Instead of accumulating a large probabilistic model rather than a Gaussian one. The proposed
IiI vector and C 1 matrix, the algorithm updates the cost by approach aims to model that, within the encoder resolution
computing the contribution at each j node as a mahalanobis range, measurements should be very likely, but out of this res­
distance between Xf and xHj. This is equivalent to equation 8 olution interval, probability should fall down rapidly. In terms
since C1 matrix is a block-built matrix, but avoids computation of a cost function, Fc(si), the proposed model results at zero
of the inverse of a large matrix. cost for values of t5b inside the encoder resolution range, but
the cost grows up when t5b is out of this resolution tolerance
B. Cable Length Constraint interval. This model is summarized with the following cost
function:
During the time period of the ithwindow, measurements
acquired by the cable encoder are the set {,A,HjITHj E
h,T i + Nw]} , so keyframes are directly choosen when new
cable encoder data is available. Cable encoder measures the
total unfolded length of the cable tethering the robot. Within
a window, we define the following difference: Other noise sources could be the non-straight deployment
of cable length, but since the proposed model computes
(17) differences within the starting point of the window and the end

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on September 08,2023 at 18:07:29 UTC from IEEE Xplore. Restrictions apply.
 D. IMU Initialization
IMU device is mainly affected by two kinds of noise
processes. One of them is considered as a Gaussian process,
which is taken into account by weighting the IMU constraint
by the inverse of a covariance matrix (equation 12). But
a second noise process is of low dynamics, and it can be
modelled as a measurement bias for accelerometers, b" and
Fig. 3. Small green segments are parts of accumulated vertical displacement
another bias for the gyros, bw. Moreover, accelerometers also
over a window of Nw = 3. Lateral displacement follows the same idea. measure the ubiquitous gravity vector, gO, that is compensated
in ground applications with the force experimented by the floor.
So three vectors have to be estimated initially, in order to use
one (equation 17), main cable curvatures are cancelled since IMU data properly, gO, b" and bw, since our approach does
they affect both cable length measurements Ai and AHNw. not observe such quantities online. Such initalization assumes
that the platform is completely stopped while running the
C. Wheel Constraint optimization loop, forcing the integrated velocities to be 0 and
the orientation quaternion to be [1 0 0 OlT. So the state to be
optimized is the 9-dimensional vector of [gO b" bw]. Initial
Wheel constraint can be applied under the assumption
that the platform is wheeled, and slippage and jumping are
guess for the gravity is the mean acceleration during a short
relatively low. It assumes the heuristics that lateral and vertical
period (for instance, one second), and initial bias are set to
displacements of the vehicle are small, so it bounds these
zero vectors. Subsection V-B reports results on runnig such
motions. For a window state point si, lateral and vertical
initialization with the actual platform used for the experiments.
displacements, <51 and <5L can be computed as:
Nw-l
<5i (si)
L,v '""'
� UL V
,
(RHj)-lvHj �H
t+Jj+l' (21) E. Initial Guess

j= O
=

When calling a minImIzation routine, an initial guess

where UL,v are the lateral and vertical axis of the vehicle has to be provided, labelled as si, which is a window
coordinate frame. Figure 3 shows the vertical displacements state point from which the optimization will start compu­
in a window of Nw 3. =
tations. For ith window, available measurements are 0/,
wt, �t , Vth E h,THNwl and AHj, Vjh+j E
The proposed model allows, up to a certain threshold, non­ h,THNwl· Moreover we know also the past estimate history,
forward displacements (lateral or vertical). So the probabilistic {xo ... xi-I ... xHNw-1}. So the initial guess for the opti­
model is neither Gaussian as it wasn't for the cable constraint mizer, provided as an initial window state point is:
discussed in subsection III-B. We use a similar approach,
which in terms of cost function is modelled as equation 20, (22)
but with parameters {L'CYL and {v,CYv .
which indicates that optimization result of the last window is
Algorith 2 summarizes how cable and wheel constraints used as an initial guess for the current ith window, poping
are computed. front Xi-1 and pushing back an IMU prediction for the last
platform state, x}+Nw.
Algorithm 2 CABLE & WHEEL Cost Function
INPUTS At the beginning, the first platform state point is set to
· WINDOW STATE POINT: si zero, xO [0 0 0 0 0 0 f, which assumes a stopped platform
=

· SENSOR DATA: {AHjITHj E h,THNw]} before operations and sets the localization coordinate reference
· KEYFRAME TIMESTAMPS: {THj,Vj 1 . . . Nw} = frame at the initial point of platform motion.
· PARAMETERS: CYc'{C'CYL){L)CYv,{v
OUTPUT: Fc(si),Fw(si) IV. IMPLEMENTATION DETAILS
gi = AHNw _
Ai
gi = 0 A software prototype has been implemented in C++ under
for j = 0; j < Nw; j + + do Ubuntu 12.04, by using the algebra and optimization classes
d = (RHj)-lvHj ��!;+l of dlib library [3]. This section details three practical issues:
gi+ = UF ' d
timeline, skipping cable data and threshold value.
<5}+ = UL ' d a) Timeline: Since IMU and encoder devices provide
<5v+ = UV ' d data asynchronously and at different rates, timeline concepts
end for are critical, and iteration indexes are based on them. Figure 4
<5b gi - gi
=
shows the continuous timeline and the different involved
Fe Fc(<5b,cyc' {c) Ilequation 20
=
indexes running over it. As suggested in figure 4, the first
FL = FL(Si,CJL'{L) Ilequation 20 platform state of the window is called anchored, since their
Fv = Fv(S�,CJv'{v) Ilequation 20 components are not part of the optimization space, but it is
Fw = FL + Fv required as a necessary starting point to integrate IMU data
return Fe, Fw
(see subsection III-A), as well as to compute expected forward
displacements (see subsection III-B).

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on September 08,2023 at 18:07:29 UTC from IEEE Xplore. Restrictions apply.
 b) Skipping Cable Data: Cable encoder measures, at
10Hz, the length of the unfolded cable tethering the robot,
with a 1em resolution error. In the other hand, IMU is a much
higher sensitive device that measures inertial data at 100Hz,
so a big difference in the nature of both measurements exists.
Since robot speeds are usually below 0.5m/ s, it has been
observed that skipping (ignoring) several consecutive cable
data points results in better optimized trajectories, computed Fig. 5. JT wheeled platform used to collect data.
faster, probably because of cable data needs amplitude to
have an important corrective effect over IMU data. Otherwise,
within a short window, cable data falls under the resolution
noise level. So a parameter Ns is set, indicating how many
consecutive cable data points are periodically skipped from the
raw cable data stream. Values for Ns will be typically between
o and 3. Higher values of Ns could cause large extrapolation
errors when computing expected forward increments (see
equation 18).
e) Threshold Values: In subsections III-B and III-C a
non-Gaussian probabilistic model has been proposed, depend­
ing on a threshold. For cable length cost function, Fe (see
equation 20), the threshold is '"Ye' which is set to 0.5em, so
algorithm allows absolute differences between expected and
measured cable increments up to 0.5em within a window.
Beyond '"Ye' cost function behaves as a mahalanobis distance
with (Je set to 1em. For wheel constraint case, threshold
'"YL '"Yv
= 1em, so the maximum slippage/ftying/falling
=
Fig. 6. On the left, a georeferenced map of the JT garden area, where the
pipe run by the robot is deployed. On the right, a picture of that pipe. Distance
tolerated by the model will be up to 1em inside a window, between a reference corner and the start of the pipe is marked with a d, and
and (JL (Jv =1mm. =
was measured to 3.1m.

V. EXPERIMENTAL RES ULTS

georeferenced map, shown on the left side of figure 6, was
A. Data Set & Ground Truth used to extract a ground truth of such pipe, so estimated
On 6th and 7th July 2012, at JT headquarters [6], in Lindau, trajectories can be compared with the reference pipe layout.
Germany, several data acquisition sessions were carried out, Tractor velocities were always below O.4m/ s.
with the aim of having a data set of a mobile/tethered platform Even if this ground truth data is metrically precise, it does
(figure 5) running inside a measured pipe, while collecting data not represent exactly the true trajectory done by the robot.
from an IMU device and a cable encoder, both data streams Instead, it represents the geometry of the pipe layout where
stored with timestamps provided by the same machine. the experiments were performed.
The data set consists on several forward and backward
trajectories, with some trials at constant velocity and other B. Gravity and Bias Estimation
trials changing speeds, but always running inside a measured
As discussed in subsection III-D, IMU integration requires
pipe, pictured on the right side of figure 6. A scaled and
to initialize the gravity vector, as well as accelerometer and
gyroscope bias. Before each test, we used a sequence of 15s
j+
windowed data: (LeiT E[ Tj, Tj Nw-1lJ, p... jv j=O N w) duration of IMU data, with the platform completely stopped
I e + . . .

I to compute these three vectors. For illustrative purposes, the

)."i+1 ",i+Nw
Ai following are the values of the IMU bias for the first trial test:

t,t-3 t,t-2 t,t-t l,t l,t+ 1 l,t+2 l,t+3 "t+ 4 l,t+5

b", = [0.0018887 - 0.0017985 - 0.0031113f m/ s2
bw [-0.0000181 0.0000641 0.0000574]T rad/ s
1 1 1.1 1 1 1 1.1
=

Gravity was also estimated in the same loop. For the case
time,'t of the same trial as above, estimated gravity initialization was:
i+l +Nw gO [-0.6064705 9.7851897 - 0.1997489]T m/ S2 (23)
xi(anchored) Xi
=

X
�I ----------------------------�I
window state: Si C. Trajectory Estimation

Fig. 4. Different iteration indexes running over timeline. Red points on

This section provides results of the estimated trajectory run
the timeline mark cable data timestamp and smail black ticks mark IMU by the robot inside the pipe shown in figure 6. Several execu­
timestamp. Sensor data used to optimize ith window is grouped on top, and tions of the optimization were run, modifying two parameters
platform states forming the window state are grouped below. in order to tune them for adequate values. The parameters to

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on September 08,2023 at 18:07:29 UTC from IEEE Xplore. Restrictions apply.
 be tunned were Ns and Nw. Ns is the number of cable data VI. CONCLUSION
skip points, being Ns 0 if all points are taken, so 0 points
=

This paper proposes a complementary technique for vi­

are skipped. The other parameter to investigate, Nw, is the
window length, being Nw 1 the reduced case where only
=
suallinertial localization of a mobile robot intended to be used
in sewer pipe network scenarios. The approach uses sensor data
one platform state is optimized. Both parameters have effects
from an IMU and from a cable encoder, the later measuring
on the computation time required to execute the estimation
loop. Current implementation runs on a standard laptop PC the length of an unfolded cable linking the mobile robot
with its starting point of operations. Data fusion is based on
(intel core is), keeping real-time peformance when Nw < 10.
the Graph SLAM framework and also requires to add extra
Figure 7 plots results obtained running four executions, setting
constraints provided by assuming low slippage and jumping
Nw 7, 15 and Ns 0, 1.
= =

of the platform, which are reasonable assumptions for wheeled

inspection platforms operating at low speeds. Optimization is
Trajectory Estimates
preformed through a set of consecutive platform states. The
.30 -F==== reported results show that the approach reduces considerably
the IMU drift, so it is well suited to be considered as a
·25
complementary technique to solve situations where visual
·20
feature tracking could fail, due to environment situations, such
as low texture areas, hard light reflections or sudden water
flows or drops.
Z[m]
·15

·10

ACKNOW LEDGMENT
-- GToondTruth
Nw�7;N.�O;
__

__ Nw=15;N,,=D;
This work is carried out under the European FP7-SME
-- Nw=7;NIJ= 1; PipeGuard project with agreement number 286580. Authors
Nw�15;N,�1; would also thank the engineering support received by JT­
'---------:-�-,-,_,,: Y(m]
__

- -;::;
: -:;:-
:;::-;::t_ .,
-10 -8 -6 -4 -2 elektronik and Inspiralia, which are also partners of the project.
X[m]
0 2

REFERENCES
Fig. 7. Estimated trajectories with different values for Nw and Ns compared
with ground truth pipe profile (black).
[1] M. Agrawal and K. Konolige. FrameSLAM: From Bundle Adjustment
to Real-Time Visual Mapping. IEEE Transactions on Robotics, 24(5),
October 2008.
[2] AJ. Davison. Real-Time Simultaneous Localisation and Mapping with
We were also interested to identify the effect of imposing a Single Camera. In Proceedingg of the International Conference on
Computer Vision (ICCV), 2003.
each constraint separately, so next results compare trajectory
estimates (figure 8) for four cases: (1) only IMU constraint, [3] dLib. dlib c++ library website. dlib.net.

(2)IMU and cable constraints, (3) IMU and wheel constraints [4] G. Grisetti, R. Kiimmerle, C. Stachniss, and W. Burgard. A tutorial on
graph-based SLAM. IEEE Transactions on Intelligent Transportation
and (3) IMU,cable and wheel constraints. All executions were Systems Magazine, 2:31-43, 2010.
done by setting Nw 7 and Ns 1. We see how, only
= =

[5] P. Hansen, H. Alismail, P. Rander, and B. Browning. Monocular visual

when all three constraints are jointly considered, estimates are odometry for robot localization in LNG pipes. In Proceedings of the
close to ground truth. In the case where cable constraint is not IEEE International Conference on Robotics and Automation (ICRA),
considered, forward vehicle velocity has no bound to grow up, Shanghai, China. May, 2011.
so an amplified trajectory arises (blue plot in figure 8). [6] IT-elektronik. Jt-elektronik website. www.jt-elektronik.de/.
[7] M.l. A. Lourakis and A.A. Argyros. SBA: A Software Package for
Generic Sparse Bundle Adjustment. ACM Trans. Math. Software,
Trajectory Estimates 36(1):1-30, 2009.

.120 t===========� [8] T. Lupton and S. Sukkarieh. Visual-Inertial-Aided Navigation for High­
Dynamic Motion in Built Environments Without Initial Conditions.
)- - - - - - � - - -
- - - -- IEEE Transactions on Robotics, 28(1), 2012.
-100 I
1 [9] M. Maimone, Y. Cheng, and L. Matthies. Two Years of Visual Odometry
'I
·80
1
on the Mars Exploration Rovers. Journal of Field Robotics, 3(24):169-
'1---- ----- 186, 2007.
1
Z[m]
·60

1 [10] A. Martinelli. Vision and IMU Data Fusion: Closed-Form Solutions

-40 I for Attitude, Speed, Absolute Scale, and Bias Determination. IEEE
1
Transactions on Robotics, February 2012.
-20
'
: ---- --- -- GToondTruth
[II] C. Mei, G. Sibley, M. Cummins, P. Newman, and I. Reid. Rslam: A
1
__ OnlylMU
o
1 -- lMU+Cable system for large-scale mapping in constant-time using stereo. Interna­
tional Journal of Computer Vision, pages 1 - 17, June 2010.
}-_______ __ lMU+Wheel
20
·50
- -- __ lMU+Cable+Wheel
[12] D. Nister, O. Naroditsky, and J. Bergen. Visual Odometry. In
8

X[m] Proceedings of the IEEE Conference on Computer Vision and Pattern

50
100

Recognition (CVPR), 2004.

[13] S. Thrun and M. Montemerlo. The GraphSLAM Algorithm With
Applications to Large-Scale Mapping of Urban Structures. International
Fig. 8. Estimated trajectories considering different constraint sets. For all
Journal of Robotics Research, 25:403-430, 2006.
cases, Nw = 7 and Ns = 1 .

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on September 08,2023 at 18:07:29 UTC from IEEE Xplore. Restrictions apply.