Enhanced SLAM For Autonomous Mobile Robots Using Unscented Kalman Filter and Neural Network
Enhanced SLAM For Autonomous Mobile Robots Using Unscented Kalman Filter and Neural Network
Enhanced SLAM For Autonomous Mobile Robots Using Unscented Kalman Filter and Neural Network
Indian Journal of Science and Technology, Vol 8(20), DOI:10.17485/ijst/2015/v8i20/52470, August 2015 ISSN (Online) : 0974-5645
2
Department of Computer and IT, Hadaf Higher Education Institute, Sari, Iran; amir.panah2020@gmail.com
3
Department of Computer, Yazd Branch, Islamic Azad University, Yazd, Iran; amin.panah2020@gmail.com
4
Mazandaran University of Medical Sciences, Sari, Iran; samere.fallahpour@gmail.com
5
Young Researchers and Elite Club, Qazvin Branch, Islamic Azad University, Qazvin, Iran
Abstract
The novel method of mobile robot Simultaneous Localization And Mapping (SLAM), which is implemented by optimized
Unscented Kalman Filter (UKF) Via a Radial Basis Function (RBF) for autonomous robot in unknown indoor environment
is proposed. For atone the Unscented Kalman Filter based SLAM errors intrinsically caused by its linearization process, the
Radial Basis Function Network is composed with Unscented Kalman Filter. A mobile robot localizes itself autonomously
and makes a map simultaneously while it is tracking in an unknown environment. The offered approach has some benefits
in handling a robotic system with nonlinear movements because of the learning feature of the Radial Basis Function. The
simulation results show the powers and effectiveness of the proposed algorithm comparing with a Standard UKF.
2. Related Algorithms for SLAM Jacobian matrix. Since, the processing noise in this system
is accumulative; therefore the augmented state vector is
2.1 Radial basis Function Neural Network used to implementation this approach. In this approach,
The idea of Radial Basis Function (RBF) Networks the mean and covariance estimation are calculated with
derives from the theory of function approximation. The considering the second order of the Taylor series.
RBF network is a three-layer feed-forward network. Input Assume that a random variable x with covariance PX
layer –input layer has one neuron for each predictor and mean μ is defined and also a random variable z as
variable. Input neurons standardizes the limited area with x is associated: z=f(x). In unscented transformation,
of the values by reduce the median and dividing by the to gain the covariance and mean random variable z,
interquartile range. Then the input neurons require sigma points that are set of weighted points are used. This
the values to each of the neurons in the hidden layer. points should be selected so that have a covariance PX and
Hidden layer –has an unlimited number of neurons that mean μ.
the optimal number is decided by the training process. Calculate the set of 2n+1 sigma points from the columns
The out coming value is transmitted to the output layer. of the matrix (n + l )Px :
Output layer – The value of a hidden layer that come to
l (1)
this layer is multiplied by a weight associated with the X0 = m, W0 =
n+ l
neuron and transmitted to the output which adds up the l
weighted values and presents this sum as the output of Xi = m + ( (n + l )Px ) , Wi =
i 2(n+ l )
(2)
the network. That generally uses a linear transfer function
l (3)
for the output units and a nonlinear transfer function Xi+n = m - ( )
(n + l )Px , Wi+n =
i 2(n+ l )
(normally the Gaussian function) for the hidden units.
The nonlinear transfer function (Gaussian function)
l = a 2 (n + b ) - n (4)
is applied to the net input to produce a radial function
of the distance. The output units implement a weighted n number of state variables are added. In the above, Wi
sum of the hidden unit outputs7. The structure of Radial is the weight which is dependent with the i-th point and k
Basis Function shows in ( Figure 1). two steps of Network also is used for adjusting the filter more correctly9.
training: first, determining the weights from the input to The UKF makes use of the unscented transform
the hidden layer; then, the weights from the hidden layer described above to give a Gaussian approximation to the
to the output layer are determined8. filtering solution of non-linear optimal filtering problems
of form, but restated here for convenience:
x k = f ( x k-1 , uk-1 , ek ) (5)
z k = h( x k , uk , dk ) (6)
Where x is state
vector and u is control input and
ε δ, are the system noise and the measurement noise,
respectively. In the first phase of implementing this filter,
the augment state vector will become as the following
form:
é Xk ù
ê ú (7)
Xk = ê e ú
Q
ê ú
êd ú
ë û
Figure 1. A Radial Basis Function Neural Network
Structure. In continue of this paper, all the formulas that are
used in the UKF in which it includes of two main from
2.2 Unscented Kalman Filter sections: Measurement update and Time update10.
UKF introduced by Uhlmann and Julier in 1997 for the • The Time Update
first time. This filter structure is based on unscented X kQ = f ( X kQ , uk , ek ) (8)
transformation. 2n
mk = å i=0 wi XiQ,k
This filter is built based on transformation as unscented (9)
2n T
transformation. In the UKF, there is no need to calculate Pk = å i=0 wi ë Xi ,k - mk ùû éë XiQ,k - mk ùû (10)
é Q
2 Vol 8 (20) | August 2015 | www.indjst.org Indian Journal of Science and Technology
Omid Panah, Amir Panah, Amin Panah and Samere Fallahpour
A Hybrid filter is proposed in this part, a RBF supplemented The state transition has the form of Eq. (21):
for acting as an observer to learn the system completely X kQ = f ( X kQ-1 , uk-1 ) + N (0, ek ) (21)
online. The mean, μk which is extracted from values (x y θ
ε δ) using the RBF algorithm, use for the prediction step, εk is the process noise, f shows the nonlinear functions,
as shown in ( Figure 2). and uk is control input.
For the Taylor development of function, f its derivative
is used with respect to X kQ , as shown in Equation (22).
¶f ( X ku-1 , uk )
f ' ( X kQ-1 , uk ) = (22)
¶X kQ
(23)
Figure 2. The Architecture of the Hybrid Filter SLAM With the exchange values gained from equations 1, 2,
3, 4, 5, previous covariance and mean have the following
Main Basic inputs are covariance, mean which are form of:
2n
computed by previous input, uk-1, and present input, uk . in mk = å w i X i,k
i=0
a
(24)
a prediction step, The robot computes the previous mean 2n
Pk = å i=0 w i[X i,k
a a
- mk ][X i,k - mk ]T (25)
and covariance and then, in observation step, it computes
a Kalman gain, present mean and covariance and defined As demonstrated in Equation (26), the observation
features. model, zk contains of the nonlinear measurement function
RBF neural network can execute as a fast and correct h and the observation noise δk .
means of approximating a nonlinear mapping based on
observed data.
Vol 8 (20) | August 2015 | www.indjst.org Indian Journal of Science and Technology 3
Enhanced SLAM for Autonomous Mobile Robots using Unscented Kalman Filter and Neural Network
é rki ù
z k = h(YkQ ) + N (0, dk ) = êê i úú =
4. Simulations
ëêÆk ûú
To showing the effectiveness of offered algorithm, we used
é 2ù
+ (mki , y - y k ) ú
2
ê (mki ,x - xk ) the amend Matlab code that was developed by Bailey12.
ê ú
ê æ mi
- y ö ú + N (0, dk ) (26) The simulation was carry out by a robot with maximum
ê -1 ç k , y
tan ç i
k÷
÷- q ú speeds of 3[m/sec] and wheel diameter of 1[m]. speed
ê çè mk ,x - x k ÷÷ø k úú
êë û and the maximum steering angle are 15[°/sec] and 25[°]
m = (mx my )
i i i T (27) respectively.
For observation step, the number of arbitrary features
2n
z k = å i=0 w i z k (28) around waypoints was used. In the observation step, a
range bearing sensor model and an observation model
were used to measure robot pose and the feature position,
3.2 The Measurement Update (Correct)
which includes 1[°] in bearing and a noise with level of
To obtain the values Pxkyk and Pxkyk , it is necessary to
Q 0.1[m] in range. The sensor range is 15[m] for small areas
compute X k μk z k zk that are compute in equations 18,
and 25[m] for large areas.
24, 26, 28, with replacement of these values. To compute
In this paper, two navigation cases of the robot
the Kalman gain Kk, we need to compute Pxkyk and Pxkyk
are surveyed: a Circular navigation, and Widespread
in the feature-based maps. we will have the following
navigation.
equations:
2n
Pxkxk = å wi [zi,k -z k ][zi,k - z k ]T (29)
i=0
4.1 Navigation on Circular Map
2n
Pxkyk =å w [ Xi,ka -mk ][zi,k - z k ]T
i=0 i (30) In circular navigation, the standard UKF based navigation
and proposed filter based navigation are shown in (Figure
K k = Pxkyk Pxkxk -1 (31) 3). Bold black line is Robot path and the dashed line,
show the paths of robots should traverse, based on data
Considering Hybrid algorithm in continuous. RBF described by the actual odometry. In (Figure 4), the bold
algorithm is included with train through input data and gray line (RBFUKF) and the dashed black line (UKF)
measurement values. In the training process, weights are are the x, y, and θ errors in the case of UKF SLAM and
determined based on the communication of input and proposed filter SLAM, and in Table.1 see Mean Square
each hidden layers. RBF require higher weight to aim Error of it.
value on the higher relations between poses and heading
angle with comparing to measurement. In addition, the
second weight, ω0,equals zero because the output offset is
zero. Therefore, new estimated mean, can be described as
in Equation (32)11 (0≤ j ≤ J − 1)
æ æ j j 2 öö
ç j-1 ç || m - d || ÷÷÷÷
j j-1 j j
( ) (
j-1 j j
m k = w0 + å j=0 wij m = x å j=0 j m
k k k k ( ))
= x ççå j=0 expçç- k j
ç 2 ÷÷
÷ø÷÷ø÷÷
èç èçç 2(t )
(32)
j
d is the center of the j-th basis function with the same
dimension of the input vector and μk is an n-dimensional
input vector. τi explain the width of the basis function, N is Figure 3. Navigation result on circular map.
the number of hidden layer’s nodes, || mkj - d j || explains the Table 1. Mean Square Error
j
Euclidean norm of showing the distance between mk and, d
j
j
mse UKFSLAM RBF UKFSLAM
and jk ( x ) means the answer of the j-th basis function of the x 0.1681 0.1119
input vector with a maximum value at d j. y 0.1934 0.0583
mk = mk + K k (z k - z k ) (33)
θ 0.0282 0.0280
P = P - K P K T (34)
k k k xkxk k
4 Vol 8 (20) | August 2015 | www.indjst.org Indian Journal of Science and Technology
Omid Panah, Amir Panah, Amin Panah and Samere Fallahpour
5. Conclusion
In this contribution, a new approach proposes UKF
SLAM based on RBF Network method for a mobile
robot. The RBFUKF SLAM on a mobile robot, to make
Figure 4. Navigation Errors on Circular Map. up for the UKF SLAM error inherently caused by its noise
assumption and linearization process. The proposed
4.2 Navigation on Widespread map algorithm contains of two steps: the UKF algorithm and
In this case, the UKF based navigation and proposed filter the RBF Neural Network. The simulation results for two
based navigation are shown in (Figure 5). The bold black different navigation cases show that the improvement of
line is Robot path and the dashed line, show the paths of the proposed filter based on RBF as compared with the
robots should traverse. (In Figure 6.), the dashed black standard UKF SLAM and it also shown that algorithms
line (UKF) and the bold gray line (RBFUKF) are the x, y, has good results in wider environment but we require to
and θ errors in the case of UKF SLAM and Hybrid filter use long range sensors. To define the robustness of the
SLAM with RBF algorithm, respectively and in Table 2. proposed algorithm, simulation in Matlab is performed.
see Mean Square Error of it. Based on the simulation results, Standard UKF SLAM has
more errors than proposed filter.
6. References
1. Bailey T. Mobile robot localisation and mapping in exten-
sive outdoor environments [degree of Doctor of Philoso-
phy]. Australian Centre for Field Robotics Department of
Aerospace, Mechanical and Mechatronic Engineering the
University of Sydney; 2002 Aug.
2. Choi K-S, Lee S-G. Enhanced SLAM for a mobile robot
using extended kalman filter and neural networks. Inter-
national Journal of Precision Engineering and Manufactur-
Figure 5. Navigation result on widespread map. ing. 2010; 11(2):255–64.
3. Frese U. A discussion of simultaneous localization and
mapping. Autonomous Robots. Netherlands: Springer Sci-
ence, Business Media, Inc; 2006.
4. Ranjan TN, Nherakkol A, Navelkar G. Navigation of au-
tonomous underwater vehicle using extended kalman filter.
Trends in Intelligent Robotics. 2010. p. 1–9.
5. Panah A, Faez K. hybrid filter based simultaneous local-
ization and mapping for a mobile robot. 16th International
Conference on Image Analysis and Processing; 2011 Sep
14; Italy.
6. Chang HH, Lin SY, Chen YC. SLAM for indoor environ-
ment using stereo vision. Second WRI Global Congress on
Intelligent Systems. 2010.
Figure 6. Navigation Errors on Widespread map 7. Durrant-Whyte H, Bailey T. Simultaneous localization and
mapping: part I. IEEE Robotics & Automation Magazine.
2006; 13(2):99–110.
Vol 8 (20) | August 2015 | www.indjst.org Indian Journal of Science and Technology 5
Enhanced SLAM for Autonomous Mobile Robots using Unscented Kalman Filter and Neural Network
8. Wolf DF, Sukhatme GS. Mobile robot simultaneous local- and mathematics School of Electronics and Computer sci-
ization and mapping in dynamic environments. Autono- ence; 2009.
mous Robots. 2005; 19:53–65. 11. Cai X, Zhang Z, Venayagamoorthy G, Wunsch G II.
9. Ip YL, Rad AB, Wong YK. Effective method for autono- Time series prediction with recurrent neural networks
mous simultaneous localization and map building in un- trained by a hybrid pso-ea algorithm. Neurocomputing.
known indoor environments. In: Kolski S, editor, Mobile 2007;70:2342–53.
Robots: Perception and Navigation; 2007 Feb. p. 704. ISBN: 12. Bailey T. Available from: http://www.personal.acfr.usyd.
3-86611-283-1. edu.au/tbailey
10. Page SF. Multiple-Object sensor Management and op
timization [PHD thesis]. Faculty of Engineering, Science
6 Vol 8 (20) | August 2015 | www.indjst.org Indian Journal of Science and Technology