IEEE International Symposium on Industrial Electronics (ISlE 2009)

Seoul Olympic Parktel, Seoul, Korea July 5-8, 2009

Bridging GPS Outages of Tightly Coupled GPS/SINS

Based on the Adaptive Track Fusion Using RBF
Neural Network
PANG Chen-peng, LID Zao-zhen
School of Aerospace and Engineering, Beijing Institute of Technology, Beijing, 100081, China
[email protected], [email protected]

Abstract- The tightly-coupled GPS/SINS integrated of measurements. In this case the navigation system is a self-
navigation system based on the Adaptive Multi-sensor Track contained system which is operated by inertial navigation
Fusion algorithm utilizes the Adaptive Track Fusion algorithm to
fuse the output of the GPS/SINS Kalman filter and GPS Kalman
system only, and the error of the system is growing rapidly
filter. In case the GPS outages occur, accuracy of the system will with time.
be degraded rapidly due to the lack of measurements of the two In order to improve the accuracy of the tightly-coupled
Kalman filters. An algorithm of bridging GPS outages using GPS/SINS system based on the adaptive track fusion during
radial basis function neural network is presented to improve the GPS outage, a new algorithm of navigation system with
accuracy of the navigation system during the GPS outages. This
method uses radial basis function neural network to predict radial basis function neural network (RBFNN) is proposed.
measurement of GPS/SINS Kalman filter during GPS outages in The output results of the integrated navigation system designed
order to ensure the regular operation of the filter, resulting in during two GPS outages intentionally introduced are discussed
reliable performance of the system. The performance of the respectively.
method proposed is examined using intentionally introduced GPS
outages and the mathematic simulation results have shown that II. INTEGRATED NAVIGATION SYSTEM BASED ON THE
the method proposed outperforms the navigation system without ADAPTIVE TRACK FUSION ALGORITHM
the neural network.
Key word: Tightly coupled GPS/SINS integrated navigation The tightly-coupled GPS/SINS integrated navigation system
Adaptive Track Fusion algorithm based on the Adaptive Track fusion algorithm'!' is designed to
GPS outages improve the accuracy of the system in case of the Kalman filter
Radial basis function of GPS/SINS having a longtime work. In the system an
independent GPS solution is developed in parallel with the
I. INTRODUCTION GPS/INS integrated solution. The estimated results of the two
solutions are defined as the local tracks respectively and then
The tightly-couples GPS/SINS integrated navigation system
integrated based on the Adaptive Track Fusion algorithm.
based on the Adaptive Multi-sensor Track Fusion algorithm is
A. DynamicModels
a system which can improve the accuracy of the open-loop
The dynamic model'" of the tightly-coupled GPS/SINS
system in case of the Kalman filter of GPS/SINS having a
integrated Kalman filter in the navigation coordinate system
longtime work. It is composed of two Kalman filter, one is a
(east-north-uranian) is denoted as follows:
tightly-coupled GPS/SINS Kalman filter, and the other is an
independent GPS Kalman filter. The estimate results of the two XGPS/SINS(t) = FGPS/SINS(t)XGPS/SINS(t) + GGPS/SINS (t)WGPS/SINS (t)
solutions are defined as the local tracks and then integrated (1)
based on the Adaptive Track Fusion algorithm. The integrated in which
navigation solution based on the Adaptive Track Fusion X GPS/SINS (t) =[({JE' ({IN' ({Ju' 8VE, 8VN, ~l' s; t>A, s; C bX , CbY , c bZ'

algorithm can not only improve the accuracy of the system, but T
crX,cry,crZ' V x ' VI" VZ' litu ' litru ]
also ensure a precise output of the system if the estimate error
(2)
of GPS/SINS Kalman filter is divergent.
is the state variable.
The two Kalman filters of tightly-couples GPS/SINS system
( f/JE , f/JN .tp; ) is the attitude error;
based on the adaptive track fusion are both dependent on the
( 8VE , 8VN , 8Vu ) is the velocity error;
pseudo-range and pseudo-range rate information calculated
( JL,8A, 8h) is the position error;
according to the ephemeris data which is received by the GPS
( cb.'( ' cby , cbZ ) is the random constant drift of gyroscope;
receiver. However, for years the challenge of GPS receiver
(c rx ,CrY ,CrZ) is the first-order Markov process of gyroscope;
outage has been found to be so difficult in case the carrier
( V x' V I' , V Z ) is the bias of accelerometer;
makes large overload maneuvering, operates in poor weather
liu is the distance deviation equivalent to the clock deviation
condition, drives around obstacle or is encountered jam. Once
ofGPS;
the satellite signal is blocked, the filters of the navigation
system can not operate regularly any more because of the loss

978-1-4244-4349-9/09/$25.00 ©2009 IEEE

971

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:40 from IEEE Xplore. Restrictions apply.
 Ii ru is the distance deviation rate equivalent to the clock Xl = [v: ,V,1 , ~ , L, It, h ]GPS is the first local track;
frequency drift of GPS. x = [V
2 e , Vn , Vu ,L, A,h] INS is the second local track;

FGPS/SINS(t) is dynamic matrix and GGPS/SINS(t) is error P, ,P2 are the error-covariance matrixes of Xl and x2 .
transition matrix, and the detailed elements of them can be Actually the statistical distance D is the quantity that
found in [2]. WGPS/SINS(t) is system noise, which is zero mean measures how close between local track and local track 2. 1
white Gaussian noise of assumed known covariance. Considering the real-time performance and high accuracy of
The observation model!" of the GPS/SINS integrated Kalman GPS, the local track 1, Xl , is regarded as the standard track. If
filter can be expressed as: D<T, it means that the distance between Xl and x2 is in an

[H p] [Vp(t)] acceptable range and it is unnecessary to do the fusion

_ [PSINS - PGPS] _
.:» - H. + V. (t) x
calculation, with 2 being the output of the navigation system;
ZGPS/SINS - •
PSINS PGPS P
XGPS/SINS
P (3) However, if D>T, it means that the error of 2 is too large to x
accept or even divergent, in this case Xl and 2 must be fused x
= H GPS/SINS X GPS/SINS + VGPS/SINS by the SF algorithm and the result of fusion will be regarded as
where ZGPS/SINS is the measurement. PSINS and PSINS are the the output of the navigation system.
pseudo-range and pseudo-range rate of SINS, PGPS and PGPS The SF algorithm [4] is described as follows:
are the pseudo-range and pseudo-range rate measured by GPS X= alxl + a 2x2 (7)
receiver. HGPS/SINS is measurement matrix and VGPS/SINS is where X is the result of fusion and the corresponding error-
measurement noise. covariance matrix is:
The dynamic model' '' of the independent GPS Kalman filter
P- -P- 2p + ( P J2
in the Conventional Terrestrial System can be described as: ( trr, J 1 trl',- P2 (8)
XGPs(t) = FGPs(t)XGPs(t) + WGPs(t) (4)
where where
X GPS (t) = [x k, Yk' Zk' dc/ock.k' Xx , )\ , Zk' dc/oCk.J is the state
variable, which is composed of the coordinate position and
a. = _1_/(_1_ + _l_J-I ,(i = 1,2)
athletic velocity of carrier as well as the distance error and the
I tr~ tr~ i-r,
(9)
distance error rate equivalent to the clock error and the clock
frequency error of GPS respectively. FGPs(t) is dynamic matrix
p_(_l +_1
trr; trP2
J-
1

and its detailed elements can be found in [3]. WGPS(t) is system

noise, which is zero mean white Gaussian noise of assumed tr( .) represent the trace of matrix.
known covariance. III. ALGORITHM OF BRIDGING GPS OUTAGES USING THE
The measurements of the Kalman filter of GPS system are RADIAL BASIS FUNCTION NEURAL NETWORK
the pseudo-ranges and pseudo-range rates of visible GPS Generally the tightly-coupled GPS/SINS based on the
satellites, which are shown as follows: adaptive track fusion can estimate the state of carrier

= [~GPS]
effectively; however, in case the signal of the satellites is
ZGPS (5) blocked or jammed, the GPS/SINS Kalman filter and GPS
PGPS
Kalman filter are unable to operate regularly because the
B. Scheme of the Integrated Navigation System Based on the measurements are missing. In this case output of the navigation
Adaptive Track Fusion system is the solution of SINS which has increased drift in a
The integrated navigation system based on the Adaptive long time, leading to a sharp decrease in estimated results of
Track Fusion includes two Kalman filters, a Kalman filter of the system. In order to solve this problem, a algorithm of
GPS/INS and an independent Kalman filter of GPS. The bridging GPS outages based on the radial basis function neural
attitude errors, velocity errors and position errors of SINS are network[5] (RBFNN) is proposed in the following.
estimated by the GPS/INS Kalman filter, whereas the GPS
A. Radial basisfunction neural network
Kalman filter estimates the real velocity and position directly.
The conventional multi-layer perceptron neural network
The velocity and position of SINS, which has been corrected
(MLP-NN) model is based on units, which computes a
by the estimates of GPS/INS filter, together with the output of
nonlinear function of the scalar product of the input vector and
GPS filter are fused at the central fusion node to obtain the
the weight vector. An alternative architecture is the RBFNN in
velocity and position of the carrier.
which the distance between the input vector and a certain
The estimates of GPS Kalman filter and the corrected
prototype vector determines the activation of a hidden unit. A
velocity and position of SINS are denoted as two local tracks.
RBFNN module, which is a three-tier feed-forward network
Statistical distance D is defined as the distance between the
with only one hidden layer, is composed of receptive units
local track and the holistic track obtained using the SF
(neurons) that act as the operator providing the information
algorithm:
about the class to which the input signal belongs. One of the
D =(Xl - X2 ) T (P; + ~)-1 ~(~ + 2~)-1(X1 - x2 ) (6)
most distinct characteristics of RBFNN is that, compared with
In which

972

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:40 from IEEE Xplore. Restrictions apply.
the normal MLP-NNs, its transformation function of hidden kq = "(wI . -xjq ) , xbl (11)
I ~ j1 I
layer is a Gaussian function with partial responsel", Due to this J

characteristic, it probably needs more neurons in a RBFNN in which WI, represents the vector of weight at hidden layer
than a feed-forward BP (Back Propagation) network to achieve related with input layer; x q is the q-th input vector; bl , is the
the same function; however, the former needs a shorter time to threshold.
train. Moreover, the RBFNN can approximate any continuous lJ
xI
function to any degree because of its characteristic of partial
response. k lJ
I
ylJ
The neuron model of RBFNN is shown in Fig.l. The x2lJ I

distance between weight vector and threshold vector, IIdistll ' is

the independent variable of the transformation function of
RBFNN, radbas. The function prototype of the transformation bl i
function is denoted as[51: , Fig.3. Input and output of a neuron in the hidden layer
radbas(n) = e- n (10) The output is:
The transformation function reaches its maximum when the
input data is zero and output of the network increases with the r,' = exp(- (k,')' ) = exp( ~~(WIJ' -x;)' Xbl ,) = expH lw l, -X'llxbl,})
decrease of distance between the weight vector and input
(12)
vector. In Fig.l , b is the threshold used to adjust the sensitivity
The input of output layer is the weighted summation of the
of neuron.
output of each neuron in hidden layer. Because of the linear
The structure of RBFNN is shown in Fig.2. It consists of an
transformation function, the output can be denoted as:
input layer, one hidden layer and an output layer. The input
layer connects the inputs to the network. The hidden layer y q = !':XW2, (13)
applies a nonlinear transformation from the input space to the i=l

hidden space. The output layer applies a linear transformation The unique architecture of RBFNN makes their trammg
from the hidden space to the output space. The transformation procedure substantially faster than the normal MLP-NNs. The
function of the hidden layer is a radial basis function which is interpretation given to the hidden units of RBFNN leads to a
generally the Gaussian function. two-stage training procedure. In the first stage, the weight
Input Radial basis neuron vector between input layer and hidden layer WI is determined
using fast unsupervised training methods that utilize only the
input data; the second stage involves the determination of the
output layer weight vector W2 . Since these weights are defined
in a linear problem (see (13)), the second stage of training is
also fast. It should be highlighted that the basis functions are
kept fixed while the weights of the output layer are
computed'",
In this study, for each GPS satellite two neural networks are
utilized in case the GPS outage occurs . The first processes the
pseudo-range received by GPS receiver at the input and
provides an estimate of the differential pseudo-range in (3).
The second network has the pseudo-range rate received by
GPS receiver at the input while the output is the corresponding
Fig.!. The structure of radial basis neuron model
differential pseudo-range rate.
RBFNNs were chosen as the basis for the new GPS/SINS
integration technique for several reasons: (i)capable of input-
output mapping, and are, therefore, well suited for mapping the
pseudo-range/pseudo-range rate to the differential pseudo-
range /pseudo-range rate; (ii)fast, specifically in situations
where there are few input and output parameters, leading to
near-real-time operation; (iii)simple to design when compared
to MLP-NNs, as the number of layers is, by definition,
constrained to two, and the number of neurons per layer is
Fig.2.The structure of RBFNN automatically increased to achieve the desired performance.
The relationship betwe en the input and output of hidden
layer is shown in Fig.3. The input of the i-th neuron of hidden B. Design ofRBFNN-based segmentedforwardpredictor
layer is denoted as: The GPS/SINS integrated navigation system based on the
adaptive track fusion utilizes the pseudo-ranges and pseudo-

973

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:40 from IEEE Xplore. Restrictions apply.
range rates received by GPS receiver for the GPS/SINS Fig.6 depicts the prediction procedure of pseudo-range
Kalman filter and GPS Kalman filter. The measurements of RBFNN. t5p(t - w: t) is the estimation of differential pseudo-
these two Kalman filters are shown in (3) and (5). In case the range in last window which was estimated by the RBFNN. It is
GPS outage occurs, the pseudo-ranges and pseudo-range rates subtracted from P INS (t - w: t), the pseudo-range of SINS, and
received by GPS receiver are unavailable and as a result the then the estimation of GPS receiver 's pseudo-range
two Kalman filters are unable to operate regularly because of A ps(t - w : t) is obtained. In the prediction procedure, the
the lack of measurements. The structure of the proposed system RBFNN has A l's(t - w : t) at the input and the estimation of
is shown in FigA. differential pseudo-range in current window t5p(t:t + w) is
Inertial INS ¢INS , VINS , PINS gained at the output.
calculation 1---...---.1
I PGPs (t w :t ) I Innut

Training RBFNN
IPlNs(t :t+ w)- PGPs (t : t+ w)~ Criterion ~
GPS Decision Target
receiv er f - t - - - - . I making
PSINS - PGPS logic : Fig.5. The update procedure of pseudo-range RBFNN
PSINS - Pcil'S Choice
of
fusion
model
GPS/INS
Kalman fiter

Fig.6. The prediction procedure of pseudo-range RBFNN

Pseudo-range rate RBFNN is similar to the pseudo-range
RBFNN, what is different is that the input and output data is
GPS Xl pseudo-range rate and differential pseudo-range rate.
L..--1.._-+I filter In order to detect the GPS outage in a method suitable for
PGI'S ,!JGPS VP near real-time applications, the data were divided into windows.
Fig.4 . The structure 0 the navigat ion system with RBFNN
As shown in FigA, when GPS satellites are available, the To accomplish this, the data set was partitioned into sections of
calculation result of SINS x,, which has been corrected by a certain length, called the window size w. When a GPS outage
output of GPS/SINS Kalman filter, and the output of GPS is detected within the search window, the entire window is
Kalman filter x, are fused in the fusion node using the Adaptive treated as if a GPS outage exists from the beginning of the
Track Fusion algorithm. In case the signal of GPS satellites is window. Once the segments of data are analyzed to determine
blocked, the RBFNN is used to estimate the measurement of whether or not a GPS outage has occurred, they are passed to
GPS/SINS Kalman filter, the differential pseudo-range and the RBFNN modules and used to present the required pseudo-
pseudo-range rate, to ensure the regular operation of it. The range/pseudo-range rate data to networks. By presenting the
GPS Kalman filter suspends its operation because there are no data to the networks in segments, instead of one point at a time,
GPS signals available. Therefore, during GPS outage, the local it is able to model a more evident trend for the differential
track x, is regarded as the output of the navigation system and pseudo-range/pseudo-range rate with respect to time.
the fusion calculation is not made until the signal of satellites In addition, depending on the window size and the outage
recovers. length, a GPS outage will often be divided into smaller
In this project, separate RBFNNs are created for both the segments. As algorithm processes the estimations of
differential pseudo-range and pseudo-range rate, resulting in differential pseudo-range/pseudo-range rate during the current
eight separate RBFNN modules in maximum. During the window, the RBFNN modules continue to predict differential
update (training) procedure, the pseudo-range / pseudo-range pseudo-range/pseudo-range rate in the next window, allowing
rate RBFNNs are trained using the measurements of GPS for the GPS/SINS Kalman filter operate continuously during
Kalman filter, the pseudo-range/ pseudo-range rate data the outage. In this study, in order to reach the proper window
received by the GPS receiver, to provide an empirical model size, the estimate precisions of the navigation system under the
relating the differential pseudo-ranges/pseudo-range rates to condition of different window sizes are discussed in detail in
the corresponding pseudo-range/pseudo-range rate. The update the simulation section.
procedure of pseudo-range RBFNN is shown in Fig.5, where w IV. SIMULATION AND ANA LYSIS OF RESULTS
is window size which will be illustrated in detail later.
The accelerometer and gyroscope data are generated by the
INS simulator. For the gyroscope, the random constant drift

974

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:40 from IEEE Xplore. Restrictions apply.
intensity is 5°Jh (10"), the first-order Markov process intensity networks require a sufficient number of data to be trained
is 0.35°Jh (10") , the Gaussian white noise intensity is properly, whereas increasing the window size too much can
O.l °Jh (10"), and the correlative time is 100s; the bias of the result in over-training of the networks, yielding larger position
accelerometer is 5mg (10") and its correlative time is 60s. errors. 100

The position, the pseudo-range and pseudo-distance rate of 90

each GPS satellite can be calculated according to the ephemeris 80

data. The GPS receiver is the C/A code receiver, of which 70

E
pseudo-range measurement error is 20m (10") and stochastic c- so
°
E
pseudo-random error rates is 0.1m/s (10") . 80
°'"
Q.

It is supposed that the time of the simulation is 300s in '0 40

length and two GPS outages are intentionally introduced to

e<;; 30

20
occur during this period. Both 18s and 36s outages are trialed
10
in order to evaluate the effectiveness of the improved system _____._1
0
under varying conditions . TABLE1 provides the start and stop
· 10
timings for these GPS outages. 80 85 90 95 100 105 110 115 120
nrnets)
TABLE I
LOCAn ONS OF GPS OUTAGESWITHRESPECT To TIME FORTHE SIMULAn ON Fig.7. Error of estimated position dur ing 18s GPS outage using the navigation
Outage Length of time for Time from start of system with out RBFNN
numb er outages(s) simulation for outages(s)
400
I 18 90 to 108
2 36 210 to 246 380

Fig.7 and Fig.8 depict the errors of estimated latitudes and 300

longitudes of carrier using the GPS/S1NS integrated navigation E 280

C-
system based on the adaptive track fusion, which is not E° 200
improved, during these two outages. It is clearly in Fig.7 and '"
8.
'0 180
Fig.8 that precision of the navigation system decreases rapidly e<;;
during the GPS outages. This is because that the GPS/SINS 100

Kalman filter and GPS Kalman filter can not operate regularly 80

during the outages for the loss of measurements, therefore, the

calculation result of SINS which has not been corrected is · 50 r-.----.~,.....--.-...,.......,,.,...__._.,.......,~__,__,.....,~,....,___,
regarded as the output of the navigation system. 200 205 210 215 220 225 230 235 240 245 250
time(s )
In generally, the window size has a cons iderable effect on
Fig.8. Error of estimated position during 36s GPS outage using the
the accuracy of the overall performance . Therefore, an analysis
navigation system without RBFNN
is decided to perform to find the best window size resulting in
the smallest root mean square (RMS) error overall all GPS Comparing of Errors For Vary ing Window
outages. For this purpose, window sizes varying in length from Sizes
20
1.5 to 18s were examined . The root mean square error for each E 18
of the latitude and longitude of the 18s GPS outage was ';j;' 16

calculated and the results are plotted in Fig.9. From Fig.9 it can g
u.J
14
t: 12
be seen that increasing the window sizes resulted in smaller :~ 10
RMS errors for each position. This is because that more data :£ 8
6
was used to train the networks and thus an improved C/l
:::E 4
positioning accuracy is gained. cr; 2 h 1111
Fig.10 examines the resulting errors for the 18s GPS outage o I I Ilih
1.5 3 4.5 6 9
in detail. It can be noticed that the latitude error is large, but Window Sizes(s)
decreases from its maximum value after 12s out of the 18s GPS
outage. Although the longitude error increases slowly during II!I Latitude 0 Longitude
the GPS outage, the increment is much smaller compared with
Fig .9. Root mean square position errors for various window sizes for 18s
the one in Fig.7.
GPS outage
Moreover, the 36s GPS outage is discussed. The same
The position errors for the 36s GPS outage are shown in
analysis for the best window size is repeated and the root mean
detail in Fig.12. The latitude error increases from the beginning
square error for each of the latitude and longitude was
of outage to approximate 220s and then decreases gradually
calculated and the results were plotted in Fig.ll.It is clear that
until the end, whereas the longitude error demonstrates a slow
the 12s window size is the most effective, as shown in Fig.ll .
and steady increase throughout the outage. Compared with the
Then, the larger window size results in the increase of RMS
position errors in Fig.8, it is evident that the navigation system
position errors. This can be explained by the fact that the neural

975

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:40 from IEEE Xplore. Restrictions apply.
with RBFNN can improve the accuracy of the system to a large lack of measurements of GPS/SlNS Kalman filter and GPS
extent. 20
Kalman filter. To solve this problem, this paper introduced a
new integration technique entitled RBF neural network. During
15 the GPS outages, measurement of the GPS/SlNS Kalman filter,
the differential pseudo-range and differential pseudo-range rate,
E 10 is predicted by the RBF neural network so that the GPS/SlNS
c:
o
E Kalman filter enable to operate regularly. Results of the
8. mathematic simulation show that the improved navigation
'0
eQ; system has better positioning accuracy than the tightly-coupled
GPS/SlNS navigation system without the RBF neural network,
·5
especially during long GPS outages.
REF ERENCES
· 10 +-~'----~'---r----"1r-----o----,~---.~---,-~---.~--,
80 85 90 95 100 105 110 115 120
[I] PANG Chen-peng and LIU Zao-zhen , "Tightly Coupled GPS/SINS
time(s ) Integrated Navigation System Based on Adaptive Track Fusion", Journal
ofSystem Simulation, 2008, 20(19):5136-5140.
Fig.IO. Error of estimated position for 18s outage using the navigation system [2] Wang Hui-nan, GPS Navigation Principles and Application, Bei ling:
with RBFNN Science Press, 2006.
[3] Shuai Ping, Chen Ding-chang, and liang Yong, "On Kalman Filtering of
GPS Navigation Calculation ", Telemetry & Telecontrol, 2001, 22(5):14-
Comparing of Errors For Varying Window
20.
80 Sizes [4] Zhang Yi, Zhou Tao-yun, and Yang Xiao-ya, "Research on filtering
g 70 algorithm based on information fusion", Electronic Measurement
Technology, 2007, 30(2): 65-67.
§ 60
[5] Ge Zhexue and Sun Zhiqiang, Neural Network Theory and Matlab
U3 50 Application. Publishing House of Electronics Industry, 2008.
'"
.g 40 [6] Lorinda Semeniuk and Aboelmagd Noureldin, "Bridging GPS outages
'en using neural network estimates of INS position and velocity errors",
r£ 30
Institute Physics Publishing, Meas Sci. Technol.l ? (2006):2783-2798 .
~ 20
e<: 10
o I l h h I
4.5 6 9 12 18
Window Sizes(s)
Latitude 0 Longitude

Fig.l l . Root mean square position errors for various window sizes for 36s
GPS outage
70

50

50

E
co 40
E
8. 30
'0
eQ; 20

· 10 -I-....--r~-.--,.....,~-,--.......,~,--.......,~,--~,...----,
200 205 210 215 220 225 230 235 240 24 5 250
nmets)

Fig.12. Error of estimated position for 36s outage using the navigation
system with RBFNN

V. CONCLUSIONS

The purpose of this research is to promote the accuracy of

tightly-coupled GPS/SlNS integrated navigation system based
on the Adaptive Track Fusion algorithm during GPS outages.
In case signals of visible GPS satellites are blocked, the
accuracy of the navigation system decreases rapidly due to the

976

Authorized licensed use limited to: ANNA UNIVERSITY. Downloaded on January 9, 2010 at 01:40 from IEEE Xplore. Restrictions apply.