A Kalman ltering approach for Integrating MEMS-based INS and GPS for

land vehicle applications

Ashish Kumar Gupta
Dept. of Electrical Engineering
Indian Institute of Technology, Madras
ee05s011@iitm.ac.in
S.S. Sarma Evani
Dept. of Electrical Engineering
Indian Institute of Technology, Madras
ee05m023@iitm.ac.in
R.David Koilpillai
Dept. of Electrical Engineering
Indian Institute of Technology, Madras
koilpillai@iitm.ac.in
Abstract
Global Positioning System (GPS) and Inertial
Navigation Systems (INS) are used for positioning and
attitude determination in a wide range of applications.
Since the usage of high performance INS is limited
by its high cost, for general application areas low-cost
INS units are used. Over the last few years, a number
of low-cost inertial sensors have become available.
However, these low cost sensors have large errors which
increase with time. So GPS measurements can be used to
correct the INS sensor errors to provide a cost-effective
solution with good accuracy in real-time navigation. The
integration of GPS and INS is done using a Kalman lter,
using the loose coupling method. The application of this
integration strategy for land vehicle navigation system
(LV NS) is discussed. In this paper, the performance
of integration system is discussed under different GPS
availability scenarios.
Keywords- Inertial navigation system, Global posi-
tioning system, Kalman lter.
1. Introduction
The rapidly expanding use of the Global Positioning Sys-
tem (GPS) enables commercial navigation devices to
be more popular and attainable for civilian applications.
GPS provides absolute positioning information covering
any part of the world during day and night [2, 7]. Al-
though GPS measurements are essential information for
our model (in general, for any navigation application),
the situation when GPS signals are unavailable (or) un-
reliable due to signal blockages, must be compensated to
provide continuous navigation solutions [3, 2]. There-
fore, presently, GPS can not be used as a stand alone
system. In order to overcome the unavailability prob-
lem in satellite-based navigation systems (GPS) and also
to be cost-effective, Micro-Electro Mechanical Systems
(MEMS) based inertial sensor technology have pushed
the development of low-cost integrated navigation sys-
tems for general applications [7]. Over the last few years,
the proliferation of low-cost inertial sensors have resulted
in an increasing interest in integrated GPS/INS systems.
Some applications require higher position update in order
of 10 to 100 Hz to track the higher dynamics of the ve-
hicles, but the commercially available GPS receivers can
not provide such higher rates due to additional process-
ing power. On the otherhand, INS, which comprises ac-
celerometers and gyroscopes, is a self-contained system
which provides accurate position, velocity and attitude
information in short intervals at very high update rates
but has time dependent error growth when operated in
stand alone mode [1]. Both the above systems are com-
plementary and hence INS can be used to bridge the gaps
in case of GPS outages. Although, these low-cost sensors
do not perform as well compared to the high-grade sen-
sors, nevertheless, when integrated with GPS the overall
perfomance of the integrated system improves. There are
many ways to couple the GPS and INS - tight coupling
and loose coupling [2, 1]. In this work we focus on sys-
tems using loosely coupled approach.
In loosely coupling mode, the position estimates from
GPS receiver and INS are used as measurements to
Kalman iter, while in tight coupling, the pseudorange
estimates from GPS are used as measurement [10]. Also
loose coupling can be done using open loop and closed
loop. In closed loop systems, the inertail error estimates
from the Kalman lter is fedback to the accelerometers
and gyros raw measurements.
In this paper, Section 2 deals with the system model
where the error state vector and its estimation proce-
dure are explained. In this Section, error modeling using
Gauss-Markov model is presented. Section 3 presents the
Kalman lter modeling for this application. Results are
shown in Section 4, and conclusions are provided in Sec-
tion 5.
2. System Model
Integration of GPS and INS using loose coupling is
shown in Fig.1. In this application, both the sys-
tems are independent of each other. The INS block,
which is shown here, consists of Inertial Measurement
Unit (IMU) and mechanization block [3, 1]. IMU
Home Sessions Authors Session 4.2 | | |
Figure 1: Loose Coupling with feedback
has 3accelerometers to measure the acceleration in
3dimensions and 3gyroscopes to measure the angular
velocities. The Mechanization block converts the IMU
measurements from body frame to required navigation
frame. Then The navigation frame measurements are in-
tegrated to obtain the velocity and the position of the ve-
hicle. Position, velocity and attitude of vehicle from INS
are denoted as r
ins
(deg) , v
ins
(m/s) and
ins
(deg) respec-
tively. Where
r
ins
=
_

ins

ins
h
ins

T
(1)
v
ins
=
_
V
Nins
V
Eins
V
Dins

T
(2)

ins
=
_
Roll Pitch Y aw

T
(3)
where
: lattitude of vehicle position
: Longitude of the vehicle position
h: Height
V
N
: North velocity
V
E
: East velocity
V
D
: Down velocity
GPS measurements can be represented as:
r
gps
=
_

gps

gps
h
gps
_
T
(4)
Errors in INS are due to biases, scale factors and mea-
surement noise. In this work, scale factor errors were not
considered. So measurement from INS system can be
modeled as
r
ins
= r +r +B
r
+
r
(5)
v
ins
= v +v +B
v
+
v
(6)

ins
= + +B

(7)
Also GPS measurements can be modeled as
r
gps
= r +
gps
(8)
where r, v and are the true information of position, ve-
locity and attitude.
r, v and are the errors in position, velocity and atti-
tude respectively.
B
r
, B
v
and B

is the bias errors in position,velocity and

attitude.

r
,
v
,

and
gps
represent measurement noises which
are assumed to be white Gaussian with zero mean.
2.1. Modeling of Biases
Biases have a serious impact on accuracy. There are two
types of biases, turn-on bias and on-run bias. Turn on
bias changes only from start up to start up while On-run
bias changes continuously. On-run biases of IMU can be
modeled as a random process. In this work, biases are
modeled as a rst order Gauss-Markov model [1], and
are assumed to be independent of each other. Estimation
and prediction of biases is done using the Kalman lter.
Using rst order Gauss-Markov model biases can be rep-
resented as [8, 4]
B
r
(k + 1) =
1

r
B
r
(k) +
r
(9)
B
v
(k + 1) =
1

v
B
v
(k) +
v
(10)
B

(k + 1) =
1

(k) +

(11)
where
r
,
v
and

represent the time-constants of

Gauss-Markov process.

r
,
v
and

represent the white Gaussian noises with

zero mean.
3. Kalman Filtering Algorithm
The heart of the integrated system is the Kalman lter
that fuses measurements from the GPS and INS sensors
to obtain an optimal estimates of the system states. This
section therefore discusses the design of the Kalman l-
ter.
3.1. State vector
The state vector of the Kalman lter consist of follow-
ing 18 elements (Six components, each with three sub-
components)
x =
_
r v B
r
B
v
B

_
T
(12)
3.2. Observation
The observations used to update the Kalman lter are the
errors between GPS and INS computed position vector,
and thus can be represented as
r
gps
r
ins
= r +B
r
+n (13)
where n is the additive white noise.
3.3. State transition
Acontinuous systemequation for INS systemcan be con-
structed as follows [4, 6]
x = Fx +Gu (14)
Where F is the dynamics matrix, x is the state vector, G
is the design matrix, and u is the forcing vector function.
The dynamics matrix F can be represented as
F =
_

_
F
ins99
0
93
0
33
0
33
0
39
(
1
r
)I
33
0
33
0
33
0
39
0
33
(
1
v
)I
33
0
33
0
39
0
33
0
33
(
1

)I
33
_

_
.
(15)
The submatrix F
ins
is taken from [1] .
I
33
is a 3 3 Identity matrix. The design matrix G can
be represented as
G =
_

_
0
33
0
33
0
33
0
33
0
33
R
n
b
0
33
0
33
0
33
0
33
0
33
R
n
b
0
33
0
33
0
33
0
33
0
33
I
33
0
33
0
33
0
33
0
33
0
33
I
33
0
33
0
33
0
33
0
33
0
33
I
33
_

_
. (16)
Where R
n
b
is Direction Cosine Matrix, used to transform
the body frame parameters into NED frame. R
n
b
depends
on the roll, pitch, yaw of the INS system. The elements
of u are the white noise whose covariance matrix is given
by
E[u(t)u()
T
] = Q(t)(t ) (17)
Q is called the spectral density matrix.
This continuous time equation can be transformed
into discrete time form [4, 6] as shown:
x
k+1
=
k
x
k
+
k
(18)
where is the state transition matrix, and
k
is the driven
response at time t = k+1 due to the presence of the input
white noise during time interval (t = k, t = k + 1) [4].
The covariance matrix associated with
k
is
E[
k

T
i
] =
_
Q
k
, i=k
0, i = k
(19)
The analytical method to nd the state transition matrix
is

k
= L
1
[(sI F)
1
] (20)
Following the numerical approximation is used

k
= exp(Ft) I +Ft (21)
3.4. Algorithm
y
k
= C
k
x
k
+
k
(22)
The Kalman gain K
k
,error covariance P
k
and estimated
state x
k
can be updated using equations given below [5,
8, 9] :
K
k
= P

k
H
T
k
(H
k
P

k
H
T
k
+R
k
)
1
(23)
x
k
= K
k
z
k
(24)
P
k
= (I K
k
H
k
)P

k
(25)
x

k+1
=
k
x
k
(26)
P

k+1
=
k
K
k

T
k
+Q
k
(27)
Where P

k
is the intermediate covariance and
y
k
= [r
k
gps
r
k
ins
] =
_
_

gps

ins

gps

ins
h
gps
h
ins
_
_
(28)
C
k
=
_
I
33
0
36
I
33
0
36
_
(29)
3.5. GPS outage
The performance of integrated system degrades during
GPS outage due to accumulation of INS errors. To im-
prove the performance, error should be predicted and es-
timated during GPS outage. Usually, the Kalman lter
estimates errors, and then, updates errors on the basis of
the present measurement. In the present work, the last
available measurement is held to update errors in Kalman
lter during GPS outage.
4. Experiments and Results
In this section, some simulation and perfomance results
are shown. Measurements from GPS are sampled at 1 Hz
while measurements from INS are sampled at 10 Hz. To-
tal duration of measurement is 35 minutes. Accelerome-
ter noise parameters are given by
a
= .017m/s/sqrt
hour in all 3 directions. Gyroscope noise parameters are
given by
g
= 0.4deg/roothour in all the 3-directions.
Initial biases for the gyros and accelerometers are given
by 7deg/hr and 10mg respectively. In Fig.2 shows the
reference trajectory and estimated trajectory for a short
segment of duration. In Fig. 3 shows the difference be-
tween GPS position and INS estimated position. It is seen
starting error is high because Kalman iter take some
time to track the GPS trajectory. In Fig. 4, covariances
of position state and error during GPS outage for 10 sec-
onds shown. States covariance reaches to zeros with time.
When there is no GPS outage performance of system is
shown in Table-1. Also stand alone INS performance is
shown in Table 1. Fig. 5 shows the trajectory when there
is only one time GPS outages for 10 seconds. It is seen
that during a GPS outage, INS trajectory deviates from
reference trajectory. During different GPS outages, per-
formance of system is shown in Table 2.
Figure 2: Navigation trajectory before and after Kalman
estimation
Table 1: RMS errors when no outage
Method Lattitude Longitude
error(m) error(m)
RMS error 0.022 0.000025
Max. error 0.2854 0.0002
Stand alone INS(rms) 289000 201500
5. Conclusions
This paper investigated the accuracy levels from a
low-cost MEMS INS when integrated with GPS for
land vehicle navigation. An 18-state Kalman lter is
used to investigate the performance of a MEMS INS in
different GPS outages. Results show that perfomance
of a 18-state system is signicantly better than stand
alone INS system. The RMS error of the stand alone
INS is approximately 290 Km in lattitude and 201
Km in longitude for 35 minutes application. Whereas
this error goes down signicantly to centimeter level
when integrated with GPS. The combined systems per-
fomance does not degrade up to 5 seconds of GPS outage.
Figure 3: Error in lattitude and Longitude when no GPS
outage
Figure 4: Covariance of Lattitude error state and Latti-
tude error when GPS outages for 10 second
Figure 5: Trajectory during GPS outage for 10 seconds
Table 2: RMS errors during GPS outage
GPS Outage Lattitude Longitude
(in sec) rms error(m) rms error(m)
1 0.089 .065
2 0.44 .32
5 0.87 0.86
10 4.59 5.4
15 10.2 10.36
Acknowledgement
We are very greateful to Dr.Ravi Swarna Babu and
Mr.Bhaskar chandrasekhar from Wipro, who provided us
MEMS-based INS and the corresponding GPS data re-
quired for this work and for their valuable guidance.
6. References
[1] D.H.Titterton and J.L.Weston. Strapdown inertial naviga-
tion technology. Peter Peregrinus Ltd., 1997.
[2] Kaplan, E. Understanding GPS principles and applica-
tions. Artech House., 1996.
[3] J. A. Farrell and M. Barth. The Global Positioning System
& Inertial Navigation. McGraw-Hill., 1999.
[4] R. G Brown, P. Y.C. Hwang. Introduction to random sig-
nals and applied Kalman ltering. John Wiley & Sons,
3
rd
edition, February 2002.
[5] Simon Haykin. Adaptive lter theory. Pearson education
4
th
edition,2002
[6] Grewal, M. and A. P. Andrews. Kalman ltering: Theory
and Practice using MATLAB. John Wiley & sons 2001
[7] Grewal, S. M., L. R. Weill and A. P. Andrews. Global
Positioning Systems, Inertial Navigation and Integration.
John Wiley & sons 2001
[8] S. M. Kay Fundamentals of Statistical Signal Processing.
Pentice Hall,Inc. 1993
[9] Lars Lindbom. Simplied Kalman Estimation of Fading
Mobile Radio Channels: high performance at LMS com-
putational load. Acoustics, Speech, and Signal Process-
ing, IEEE International Conference Vol. 3,1993, pp.352-
355
[10] Salychev, O., V.V. Voronov, M. E. Cannon, R. A. Nayak
and G. Lachapelle. Low Cost INS/GPS integration: Con-
cepts and Testing. ION NTM., 2000.