Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011


Abstract:
In this paper we present some advanced methods of obstacle
avoidance for a laser based intelligent wheelchair using
optimized Bayesian neural networks. Three autonomous
tasks have been developed for different types of
environments to improve the performance of the overall
system. The accurate usable accessible space is determined
by including the actual wheelchair dimensions in a real-time
map used as inputs to each networks. The system combines
local environmental information gathered using a laser range
finder sensor with the users intentions to select the most
suitable autonomous task in different situations. Bayesian
frame work is used to determine the optimal neural network
structure in each case. Experimental results show significant
performance improvements compared to our previously
reported shared control methods.
I. INTRODUCTION
Obstacle avoidance is one of the most fundamental tasks of
autonomous systems. Many algorithms for enabling
autonomous tasks have been developed. The most popular
being, global map, occupancy grid [1], virtual force field and
vector field histogram [2, 3]. However, most of these
algorithms have difficulty successfully operating in dense
and dynamic environments and providing smooth
trajectories and stability required by any assistive wheelchair
systems. Our Bayesian learning general obstacle avoidance
neural network was first introduced in [4]. Initial results
were encouraging but also showed that a single neural
network could not provide the desired performance in all
situations and hence further development was required.
In this paper, we present a more advanced obstacle
avoidance techniques that utilises separate neural networks
for specified tasks. The obstacle avoidance task is divided
into three sub-tasks: passing through a door, corridor and
wall following and general obstacle avoidance. This enables
the network to respond to the particular features of each task,
improving performance. Specific data acquisitions are
performed to collect the patterns used to design the neural
network for each task. Bayesian framework is then applied
to determine the optimal network structures. The training
patterns are then used in conjunction with the Bayesian
training process to improve the generalisation and
performances of each network. Our method was able to
successfully accomplish difficult navigation tasks smoothly
following a near optimum trajectory.
In addition, an assistive wheelchair system that utilises an
adaptive shared control strategy based on the Bayesian
recursive technique to select which autonomous task to use
in different situations. As the wheelchair moves it takes into
account both the users intentions and local environment
data to estimate each tasks evidence. The task is chosen
being the one with the highest evidence value.
This paper will be presented in a number of sections. The
next section reviews the Bayesian framework. The following
two sections discuss the obstacle avoidance method, the
shared control strategy based on the Bayesian recursive
technique and some experimental results. In the last section
we present our discussion and conclusions.
II. BAYESIAN NEURAL NETWORK
Bayesian neural networks first introduced by MacKay [5,
6] have the following main advantages compared to a
standard feed-forward neural network:
- The Bayesian framework automatically constrains the
weight set to optimal values for the best generalisation
during training. While a test set is used to test a
network s performance, a separate validation set in not
required, making additional data available for training.
- Different local minima of training and network
structures, with different numbers of hidden nodes, can
be compared and ranked.
A. Regularisation
In network training, using simple sum square of error as a
performance function can not prevent the over-growth of
networks weights. Large weight values can produce poor
performance with new data. An appropriate regulation
strategy can be achieved by adding a weight decay term to
the error function, E
D
, to penalise the growth of weights as
follow:

=
+ =
G
i
w i D
i
E E w S
1
) ( o |
, (1)
where S(w) is the performance function, and
i
are known
as the hyper-parameters for training and weight sets, E
Wi
is
the weight values for the i
th
group of weights and biases, and
G is the number of groups of weights and biases in the
neural network. Minimising this performance function, S(w),





Obstacle Avoidance Methods for a Laser Based Intelligent Wheelchair
Dr Triu Tuyn Hong
1
Dr. Nguyn Hng Quang
2
1
Cng ty c phn vin thng qun i Vietel
2
B mn t ng ha x nghip, Trng HBK H Ni
1
e-MAIL: trieutuyenhoang@yahoo.com
2
e-MAIL: quangnh@mail.hut.edu.vn

250
Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011

by appropriately adjusting the hyper-parameters can
converge the weight set to a global minimum ensuring both
a networks performance and generalisation.
B. Bayesian Framework
The Bayesian framework for a multi-layer perceptron
neural network is based on a Gaussian approximation. It
automatically adjusts the hyper-parameters to the most
probable value given by the training data set during the
Bayesian learning process. Different networks with different
structures and trained weight sets can be compared and
ranked to find the most suitable network for an application.
According to the Bayesian inference, the posterior
probability of the network parameters, weight set - w, of a
neural network, H, given by a training data set, D, could be
estimated by:
}}
= | o | o | o d d ) D | , ( p ) D , , | w ( p ) D | w ( p
. (2)
With a Gaussian approximation for posterior distribution
of hyper-parameters, p(, |D), this integration can
estimated as
}}
| o | o | o d d ) D | , ( p ) D , , , w ( p ) D | w ( p
MP MP
(3)
which can be simplified to
) D , , , w ( p
MP MP
| o
by using
1 ) | , ( =
}}
| o | o d d D p
as a normalization factor [7].
This mean that the most probable values
MP
,
MP
shall
maximise the posterior probability of weights. These values,

MP
,
MP
, can be estimated from their posterior of
distribution as equation follows, [9].
) , | D ( p
) D ( p
) , ( p ) , | D ( p
) D | , ( p | o
| o | o
| o =
(4)
The term p(D|, ) is called evidence of the hyper-
parameters. The log of this evidence could be estimated by
equation bellows, [9]:
) 2 ln(
2
ln
2
ln
2
ln
2
1
) ( ) , | ( ln t | o | o
N N W
A w S D p
MP
+ + =

where A is the Hessian matrix of the cost function, A = C +
B, C E
w
= VV , B E
D
= VV . The term W is the number of
network parameters, N is the number of training patterns and
w
MP
is the most probable value of weight.
The most probable values of hyper-parameters
MP
,
MP

can be estimated by equation above as:
MP
W
MP
E 2

o =
,
MP
D
MP
E
N
2

|

=
,

=
+
=
W
i i
i
1
o

(5)
where
i
is the eigenvalue of the Hessian matrix A. These
values are re-estimated during training to constrain the over
growth of weight values to ensure the generalization of the
neural network.
Bayesian framework can compare and rank different
neural networks with different structures and weight values
by estimating the probabilities of these networks. The
Bayesian formula for a network, H
i
, and its probability given
by the training data, D, is
) H | D ( p
) D ( p
) H ( p ) H | D ( p
) D | H ( p
i
i i
i
=
. (6)
The prior probability of a network is assumed to be the
same for all models and the term p(D) is independent on the
model. Hence, the posterior probability of the model can be
determined by evidence p(D|H
i
). The evidence of the model
can be calculated by estimating the integration below over
the set of network parameters - w,
}
= dw H w p H w D p H D p
i i i
) , ( ) , | ( ) | (
. (7)
Bishop evaluated the log evidence of model, H
i
, rather
than the evidence itself [9] as:
+ + =
MP
MP
D MP
MP
W MP i
W
A E E H D p o | o ln
2
| | ln
2
1
) | ( ln


|

+ + + + +
N
M M
N
MP
2
ln
2
1 2
ln
2
1
ln 2 ! ln ln
2
(8)
The different network structures are compared by estimating
the evidence by the above equation. The optimal network is
the one that has the highest evidence.
III. OBSTACLE AVOIDANCE METHOD
A. Data Acquisition
Our neural networks use usable accessible space data as
an input and providing values of steering and velocity as
outputs. A method reported in [8] for accurately determining
collision free accessible space, by combining information
from a laser range finder, the actual wheelchair dimensions
and encoder sensors mounted on the wheelchair is used to
produce a real-time map as shown in fig. 1.
The wheelchair is required to follow a number of
predetermined paths to gather data for training. These paths
are selected by the designer to simulate the previously
mentioned tasks. The movements of the wheelchair are
measured and formed as training patterns for each obstacle
avoidance sub-task.
B. Bayesian Training
The Bayesian framework is first applied to determine the
most suitable structure of a neural network for each task by
estimating the evidence of a set of neural networks with
different hidden nodes by equation 8. The collected patterns
are divided to two sets: training and testing sets. The aim of
using a testing set is to verify generalisation of these
networks. Second, all available patterns are used in training
this network under the Bayesian rule to find the most
probable weight set that improves the networks
performances and generalization. The trained networks are
then used to control the wheelchair in real-time.
251
Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011


IV. ADAPTIVE SHARED CONTROL STRATEGY BASED ON THE
BAYESIAN RECURSIVE TECHNIQUE
A. Bayesian recursive for wheelchairs shared strategy
The information in regard to the local environment and
the users intentions is continuously updated. The raw laser
data only provides the position of surrounded obstacles. This
data needs to be further refined to determine the most
suitable autonomous task. In particular, the system needs to
use this data to identify environment features such as
corridors, walls and doors in addition to general obstacles
that have to be avoided. Hence a model of environments has
to be produced.
Let us assume that the environment model is already
available and at time t the wheelchair controller has
observed a stream of information, I, as follows.
I = {U
1
, D
2
, U
2
, D
3
, , U
t-1
, D
t
} (1)
Where
U
t
: the user command at time t,
D
t
: the environment model information at time t.
To estimate the current state of the wheelchair the Markov
assumption is used, fig. 1, so that the values in any state are
only influenced by the values of the state that directly
preceded it. This assumption significantly reduces the
computational complexity by allowing the removal of
previous states that no longer effect the current state.

To select the appropriate task at time t, H
t
, the posteriors
of these autonomous tasks, called evidences, have to be
estimated based on the information of the environments and
the users intentions, D
t
and U
t-1
respectively, as follow
) D , U ,..., D , U , D , U | H ( P ) H ( Evi
t 1 t 3 2 2 1 t t
=
. (2)
Using the Bayesian theorem to take into account the
environment information:
}


=
t 1 t 2 1 t 1 t 2 1 t t
1 t 2 1 t 1 t 2 1 t t
t
dH ) U ,..., D , U | H ( P ) U ,..., D , U , H | D ( P
) U ,..., D , U | H ( P ) U ,..., D , U , H | D ( P
) H ( Evi
. (3)
Based on the Markov assumption,
}

=
t 1 t 2 1 t 1 t 2 1 t t
1 t 2 1 t t t
t
dH ) U ,..., D , U | H ( P ) U ,..., D , U , H | D ( P
) U ,..., D , U | H ( P ) H | D ( P
) H ( Evi
, (4)
) U ,..., D , U , D , U | H ( P ) H | D ( P ) H ( Evi
1 t 3 2 2 1 t t t t t
=q
,
Where
}

=
t 1 t 3 2 2 1 t 1 t 3 2 2 1 t t t
dH ) U ,..., D , U , D , U | H ( P ) U ,..., D , U , D , U , H | D ( P q
.
Take into account the users intentions, the evidences are
}

= ) H , U ,..., D , U , D , U | H ( P ) U ,..., D , U , D , U | H ( P
1 t 1 t 3 2 2 1 t 1 t 3 2 2 1 t

1 t 1 t 2 1 1 t
dH ) U ,..., D , U | H ( P

. (5)
From equations 4 and 5, we have
}

= ) H , U ,..., D , U , D , U | H ( P ) H | D ( P ) H ( Evi
1 t 1 t 3 2 2 1 t t t t t
q

1 t 1 t 2 1 1 t
dH ) U ,..., D , U | H ( P

,
}

=
1 t 1 t 2 1 1 t 1 t 1 t t t t t t
dH ) U ,..., D , U | H ( P ) H , U | H ( P ) H | D ( P ) H ( Evi q
}

=
1 t 1 t 1 t 1 t t t t t t
dH ) H ( Evi ) H , U | H ( P ) H | D ( P ) H ( Evi q
. (6)
In our discrete application with only three operating
modes, the evidences for these tasks are simplified as:
) H ( Evi ) 1 Ht , 1 Ut | H ( P ) H | D ( P ) H ( Evi
1 t
i
} DP , CWF , GOA { i
t
i t t t t

=
=

q
(7)
Where

=

=
} DP ; CWF ; GOA { i
1 t 2 1
t
i 1 t 2 1
t
i t t
) U ,..., D , U | H ( P ) U ,..., D , U , H | D ( P q
(8)
The appropriate task being selected based on its
evidences value.
B. Shared control strategy framework
The framework for the adaptive shared control strategy is
presented in the fig. 2. Initially, at t = 0 as there is no prior
environment information nor users intentions, the priors are
assumed to be equal for all autonomous tasks. When new
information arrives, the initial evidences are calculated using
the environment model data, equation 4. The final values are
then determined after considering the latest user intentions,
equation 7. Finally the most suitable autonomous task is
selected as being the one with the highest evidence.
As previously mentioned some wheelchair applications
have used offline environmental data, such as the use of a
topology map in Navchair [4]. This however means that the
wheelchair can only operate in a known environment. In
this paper we introduce a novel method of using laser
images to classify the local environment. The locations of
corridors, walls, doors and general obstacles are determined
from the raw laser data. Again we apply the Bayesian neural
network framework, as in [5, 6], as follows:
- Step 1: collect the training patterns from various
environments.
- Step 2: apply the Bayesian framework to determine the
optimal neural network structure and to train this
network for better generalization.

D
t-1

H
t-1
U
t-1

D
t

H
t
U
t

D
t+1

H
t+1

Fig. 1. The Markov chain for the wheelchair application. H is the
wheelchair state. U is the users action. D presents environment data.


Fig. 1. Useable accessible space identified as the shaded area and
real-time obstacle map (outer line) for the wheelchair.

252
Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011

V. EXPERIMENTAL RESULTS
A. General Obstacle Avoidance (GOA)
The wheelchair was made to follow a number of paths
classified as general obstacle avoidance situations to collect
data. The number of patterns gathered was 3951. This set
was divided to two sets: training and test (2374 and 1577
patterns respectively) based on the independent data
collected from the different paths.
First, a Bayesian framework was applied to determine the
most suitable structure for the general obstacle avoidance
neural network. The training results are shown in fig. 2. The
network with four hidden nodes produced the highest
evidence. This network also produced lower errors
compared to other networks when tested by the test set.
Hence this structure is the most suitable for the general
obstacle avoidance task.
Second, the data from both the training and test sets was
used to train a network with four hidden nodes applying the
Bayesian rule. During training, the Bayesian framework
constrains the growth of weights to the most probable values
by automatically adjusting the hyper-parameters, and .
After training the network was used enable the wheelchair to
perform general obstacle avoidance tasks.

In the first experiment (results shown in fig. 3) the
wheelchair was asked to access to a narrow dead-end
corridor. Our method was compared to the well-known VFH
[5], Vector Field Histogram. The VFH algorithm utilises a
polar-histogram of range data to keep to the middle of the
available free-space determined by a constant threshold. As
shown in the figure our neural network method produced a
superior result providing a smooth, stable and reliable
trajectory as the wheelchair navigated the requested path.
Conversely, the VHF algorithm was not as smooth and
guided the wheelchair extremely close to the obstacle on the
left hand side when negotiating the corner.
B. Corridor and Wall Following (C-WF)
The data acquisition procedure as described for the
general obstacle avoidance task was again used for this task.
The number of collected patterns was 3478. These were
divided into separate training and test sets (2283 and 1195
patterns for each set respectively).
The Bayesian framework was again applied to find the
most suitable network structure for this task. The training
results shown in fig. 4 found that a network with only two
hidden nodes produced the highest evidence and recorded
consistently low errors when tested by the test set. Hence
this is the most suitable network structure for the corridor/
wall following task. The most probable weight values were
obtained by applying all patterns to the Bayesian training.
The trained network then was used to control the wheelchair
when performing corridor and wall following tasks.

Fig. 4. C-WF tasks training result. The network with only two
hidden nodes can provide the highest evidence and low test errors.
The wheelchair was required to travel along the longest
wall in our laboratory as shown in fig. 5. Again the
performance of our neural network method was compared to
the VFH algorithm. Our method guided the wheelchair
smoothly and reasonably directly along the wall, moving
only slightly away from the wall where the wider free-space
was encountered. Conversely, the VFH algorithm produced

Fig. 3. GOA neural network performances.


Fig. 2. GOA tasks training result. The network with four hidden
nodes is the most suitable also providing low testing errors.

253
Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011

a less satisfactory result producing a fluctuating and less
direct path during the experiment.

C. Door Passing (DP)
Of the three sub-tasks passing through a doorway is the
most difficult. 3070 patterns were acquired during the day to
requisition process, which were divided into training and test
sub-set with 2356 and 714 patterns respectively.
After applying Bayesian training, a network with five
hidden nodes was found to produce the highest evidence,
fig.6 and small errors when tested by the test set. The most
probable weight values were obtained by applying all
patterns to the Bayesian training of the five hidden node
network. After training this neural network was used to
control the wheelchair for door passing tasks.
To test the network performance the wheelchair was
requested to travel along a corridor and then turn through a
doorway, fig. 7. The experiment found our network had
superior performance compared to the VFH algorithm. This
network made a substantial turn as the wheelchair
approached the doorway then reduced the wheelchairs
velocity at point 1 in preparation for the door passing task.
The wheelchair followed a smooth and near optimal path in
negotiating the doorway finishing the task at point 2, after
which the wheelchair moved towards the middle of the
available free-space. Conversely, the VFH algorithm was
unable to successfully navigate through the doorway,
colliding with the door frame before stopping at point 3.

D. Environmental model training results
Two experiments were designed to compare the
performance of the new shared control strategy to the
previously reported cost function shared method [4]. In the
first, the wheelchair moved through three distinct
environments, while in the second the environment included
a number of possible paths. In both experiments data was
recorded at 100 milliseconds intervals (computation circles).

0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.2
0.4
0.6
0.8
1
Computation circles
E
v
i d
e
n
c
e
s


GOA task
CWF task
DP task
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.2
0.4
0.6
0.8
1
Computation circles
E
v
i d
e
n
c
e
s

o
f

o
p
e
r
a
t
i n
g

m
o
d
e
s


GOA task
CWF task
DP task

(a) The evidence values at the wheelchair start point (point A)
8 9 10 11 12 13
0
0.2
0.4
0.6
0.8
1
Computation circles
E
v
i d
e
n
c
e
s

o
f

o
p
e
r
a
t
i n
g

m
o
d
e
s


GOA mode
CWF mode
DP mode
8 9 10 11 12 13
0
0.2
0.4
0.6
0.8
1
Computation circles
E
v
i d
e
n
c
e
s

o
f

o
p
e
r
a
t
i n
g

m
o
d
e
s


GOA mode
CWF mode
DP mode

(b) The evidence values for the first left-turn command (point1)
29 30 31 32 33 34 35 36
0
0.2
0.4
0.6
0.8
1
Computation circles
E
v
i d
e
n
c
e
s

o
f

o
p
e
r
a
t
i n
g

m
o
d
e
s


GOA mode
CWF mode
DP mode
29 30 31 32 33 34 35 36
0
0.2
0.4
0.6
0.8
1
Computation circles
E
v
i d
e
n
c
e
s

o
f

o
p
e
r
a
t
i n
g

m
o
d
e
s


GOA mode
CWF mode
DP mode

(c) Evidence values when the wheelchair turns to the door (point 2, 3)
Fig. 8: Evidence values calculated by the recursive Bayesian technique.
The results for the full Bayesian shared control framework are shown
on the right figures, while those without considering the users
intentions are shown on the left.

Fig. 7. Our network allowed the wheelchair to navigate safely and
relatively directly through the doorway. In comparison the VHF
algorithm crashed the wheelchair into the door frame.


Fig. 5. The corridor and wall following experiment showed our neural
network provided superior performance compared to VFH algorithm.


Fig. 6. Passing through a door tasks training result. The network
with five hidden nodes produced the highest evidence and provided
low error when tested by the testing set.
254
Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011


1
A
B
2

Fig. 9. Comparative performance of the adaptive shared control strategy
and cost function shared method for the first experiment.


The evidence values for three autonomous tasks are
shown in fig. 8 while the comparative results for the first
experiment are shown in fig. 9. The wheelchair started from
point A in an environment that contained only general
obstacles. Figure 8a shows that the evidence increases for
the general obstacle avoidance mode while decreasing for
the two other modes. Hence the wheelchair chooses to
operate in the general obstacle avoidance mode.
This mode of operation is maintained until the wheelchair
reaches point 1 where it detects a corridor and receives a left
turn command from the user. Figure 8b shows the adaptive
shared strategys reaction to this situation with the corridor
following mode evidence increasing and the evidence of the
other two modes decreasing. Hence the wheelchair switches
to run in the corridor following mode.
Figure 4c (left) shows the evidence for the door passing
mode increasing at point 2 in fig. 9 confirming the systems
detection of the approaching doorway. However, with no
change in intention from the user fig 8c-right (computation
circle 30 to 34) shows that the overall value of evidence for
this mode only increases a small amount to be slightly
higher than the evidence for the general obstacle avoidance
mode but significantly less than the evidence for the corridor
following mode, hence at this point the wheelchair remains
in the corridor following mode.
At point 3 (fig 9) the user indicates their intention to turn
left towards the doorway. At this point (computation circle
35 in fig.8c-right) the evidence for the door passing mode
increases to the highest value of all three modes. The
wheelchair switches to the door passing mode and
confidently navigates through the doorway.
VI. DISCUSSION AND CONCLUSION
The results suggest that Bayesian neural networks have
significant potential to solve the obstacle avoidance tasks
required by an intelligent wheelchair system. Improved
performance is achieved by dividing the overall obstacle
avoidance task into a number of sub-tasks each controlled by
using the specifically designed neural networks. In addition,
as the Bayesian framework resists overgrowth of network
weights, it promotes network generalization, assisting it to
deal with new environments. After training the networks
showed the potential to provide satisfactory real-time
performance. The optimal method of effectively combining
these networks to achieve the desired performance is the
focus of ongoing research.
The reported performance of the adaptive shared control
strategy suggests that the Bayesian recursive technique may
provide an effective solution to designing a shared control
autonomous system. It proved successful in combining the
obstacle awareness provided by the laser sensor with the
intentions of the user to provide near optimum navigation
without prior knowledge of the environment. The overall
performance of the wheelchair system was improved
significantly by correctly matching the selected autonomous
task to a given situation.
REFERENCES
[1] A. Elfes, "Using occupancy grids for mobile robot perception and
navigation," Computer, vol. 22, pp. 46 - 57, 1989.
[2] J. Borenstein and Y. Koren, "Real-time obstacle avoidance for fast
mobile robots in cluttered environments," IEEE International
Conference on Robotics and Automation, vol. 1, pp. 572 - 577,
1990.
[3] J. Borenstein and Y. Koren, "The vector field histogram-fast
obstacle avoidance for mobile robots," IEEE Transactions on
Robotics and Automation, vol. 7, pp. 278 - 288, 1991.
[4] H. T. Trieu, H. T. Nguyen, and K. Willey, "Obstacle Avoidance for
Power Wheelchair Using Bayesian Neural Network," 29th Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society, pp. 4771 - 4774, 2007.
[5] D. J. C. MacKay, "Bayesian interpolation," Neural Computation,
vol. 4, pp. 415447, 1992a.

(a) The performance of the cost function method.

A
B
C
D
Forward Command
Right Command
Right Command
(No command)
Doorway

(b) The performance of the adaptive Bayesian shared strategy.
Fig. 10. Comparative performance of the adaptive shared control
strategy (b) and cost function method (a) in the second experiment.

255
Hi ngh ton quc v iu khin v T ng ho - VCCA-2011

VCCA_2011

[6] D. J. C. MacKay, "Bayesian neural networks and density
networks," Nuclear Instruments and Methods in Physics Research,
pp. 73-80, 1995.
[7] C. M. Bishop, Neural networks for pattern recognition. Oxford:
Oxford University Press, 1995.
[8] H. T. Trieu, H. T. Nguyen, and K. Willey, "Shared Control
Strategies for Obstacle Avoidance Tasks in an Intelligent
Wheelchair," 30th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (submitted), 2008.
[9] R. C. Simpson and S. P. Levine, "Automatic adaptation in the
NavChair Assistive Wheelchair Navigation System,"
Rehabilitation Engineering, IEEE Transactions on [see also IEEE
Trans. on Neural Systems and Rehabilitation], vol. 7, pp. 452 - 463
1999.


256