ADAPTIVE COMPENSATION SIGNAL FOR A WHEELCHAIR

CONTROL USING ANFIS MODEL

Anouboudem Kinfack Fabrice Prosper, Djouani Karim and Hamam Yskandar

French South African Institue of Technology, Department of Electrical Engineering, Tshwane University of Technology

Staatsartillerie Road, Pretoria, South Africa

Keywords:

Neuro-Fuzzy inference, Bayes net, Human behaviour modelling, Autonomous vehicle navigation.

Abstract:

Wheelchairs users still face challenges when driving their standard design based vehicle. Given the matter, this

work aims to implement an assisted control for a wheelchair, depending on the driving behaviour of the user.

Therefore, a Bayesian network model will be implemented to help infer on the human behaviour. Thereafter,

the inferred state of the user will serve as input to an ANFIS model. The role of the ANFIS model is to

generate an assistive signal in order to compensate the input from the user.

1 INTRODUCTION

Assistive control aims to provide enhancements to

the technology in order to reduce people’s handi-

cap by enabling them to perform tasks they were

not able to achieve formerly. As an example, the

motion in the environment has been facilitated by

the use of wheelchairs. However, the use of stan-

dard wheelchairs brings up the issue of handling,

as users suffer from different neuromuscular impair-

ments. Therefore, there is a need to come up with a

design that will be adapted to a speciﬁc user. This

feature requires the vehicle to recognize the opera-

tion initiated by the user. In other terms, the device

must learn the driving behaviour of the user. Then,

the recognized human behaviour will help a naviga-

tion system to generate an assistive signal in order to

compensate the handicap of the user.

Even though several approaches have been imple-

mented to achieve the objective of modelling the hu-

man behaviour, very address the case of wheelchair

driving activity. Therefore, this work reviews studies

conducted in the domain of human behaviour mod-

elling on car driving activity.

The optimal control approach (Burnham et al., 1974)

considers the human as an optimal controller and

tries to identify some parameters liable to inﬂuence

the driver behaviour; however, the authors admit that

too many assumptions were made to come up with

the proposed model. The closed loop based model

(Thakur,1997) includes the mechanical time response

(from the vehicle) and the human time response which

is assumed to be a function of a daydream factor to de-

termine.

Although the cognitive architecture (Dario Salvucci

and Liu, 2001) seems to be more realistic, the model

is difﬁcult to implement as it requires the measure-

ment of some internal states of the human operator

such as the HRV (Heart Rate Variability).

In the other hand, many other studies have consid-

ered a model based on external observations. Among

others, we can mention polynomial models (Koashi

et al., 2003; Kim et al., 2004), data clustering ap-

proach (Suzuki et al., 2005), Hidden Markov Models

(Kuge et al., 1998) and Bayesian Networks (Bouslimi

et al., 2005). However, the comparison of Hidden

Markov Model and Bayes net model (Tezuka et al.,

2006) revealed that Bayesian network provided bet-

ter inference even though HMM detect the change of

operation earlier.

The method referred to as potential ﬁeld (Koren

and Boreinstein, 1991; Masoud, 2002) assumes the

presence of different forces applied to the vehicle: re-

pulsive forces are generated by obstacles while the

target point applies an attractive force to the vehi-

cle. The magnitude of the repulsive force is propor-

tional to certainty value (how conﬁdent the algorithm

is in the presence of an obstacle) and inversely propor-

tional to the distance between the centre of the robot

and the obstacle.

However, KOREN (Koren and Boreinstein, 1991)

identiﬁes four major issues associated in the applica-

tion of the method:

• Trap situations due to local minima

123

Anouboudem Kinfack Fabrice P., Yskandar H. and Karim D..

ADAPTIVE COMPENSATION SIGNAL FOR A WHEELCHAIR CONTROL USING ANFIS MODEL.

DOI: 10.5220/0003537801230129

In Proceedings of the 8th International Conference on Informatics in Control, Automation and Robotics (ICINCO-2011), pages 123-129

ISBN: 978-989-8425-74-4

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

• No passage between two close obstacles.

• Disturbances of the obstacle cause oscillations

• Oscillations in narrow passages

EMAM’s work (EMAM, 2010) on wheelchair driver

behaviour also represents the inﬂuence of obstacles

on the driver by similar forces (mostly repulsive), ex-

cept that the attractive force depends on the driver.

Another reference (Chakravarthy and Ghose, 1998)

proposes the concept of collision cone, for collision

avoidance situations. Given a vehicle and an obsta-

cle, the authors ﬁnd the locus of directions of the ve-

hicle for which there may be a collision. The collision

cone is found to be a function of the velocities of the

vehicle and the obstacle, their direction, the distance

separating them and also their shape. Therefore, in

case of possible collision (when the vehicle’s velocity

vector lies within the collision cone), the speed and

direction can be modiﬁed to bring the velocity vector

out of the collision cone. The strategy to adopt de-

pends on the time and kinematics constraints of the

vehicle. The same concept is has also been adopted

in a multi-obstacle environment (Fiorini and Shiller,

1998).

Langer et al. (Langer et al., 1994) proposed a method

for a behavioural response based on the DAMN ar-

chitecture framework developed by J. K. Rosenblatt

(Rosenblatt, 1997); the architecture is comprised of

three parts:

• The perception module processes the images

taken of the surrounding environment and iden-

tiﬁes traversable regions of the terrain.

• The management module generates a map repre-

senting the surrounding terrain.

• The planning module issues the commands sig-

nals.

All the actions of the architecture are working asyn-

chronously, sending their outputs to an arbiter which

evaluates the prevailing one. Unfortunately, the au-

thors (Langer et al., 1994) found the inability to deal

with close ends as the most important and common

failure when implementing such an architecture; this

is because the system only deals with local represen-

tation of the terrain. Moreover, the speed of the ve-

hicle is limited by the maximum range of the sensors

and by the image acquisition rate. In addition to the

latency for the full acquisition of an image, the map

generation also contributes to a non real time archi-

tecture.

Another behaviour based approach (Yinka-Banjo,

2010) uses the Bayesian network model for inference.

First, the robot is guided through an environment with

obstacles. A database of distance to obstacles and ma-

noeuvres of the (human) operator is updated; then, a

Bayesian network is built from the data gathered. This

network will serve as required knowledge of the robot

for obstacle avoidance behaviour.

NEFTI et al. (Nefti et al., 2001) proposed an ANFIS

model where they deﬁne three main tasks, each one

representinga module that implements an ANFIS sys-

tem, with sensors information as input and orientation

of the robot as output. A fusion strategy implemented

by a neural network selects the most appropriate out-

put, given the information from the sensors.

A reference (Ayari and Chatti, 2007) deﬁned two ma-

jor tasks to be executed by the robot whilst another

one (Shimizuhira et al., 2004) deﬁned three. Even

though ARAGHI and MENHAJ (Araghi and Menhaj,

2008) adopted a different approach using the concept

of collision cones, all three studies implement a fuzzy

controller serving as the decision making module.

Given the comparison of Bayesian net and Hidden

Markov Model (Tezuka et al., 2006), even though the

speed of state detection is an important aspect, this

work favours the accuracy of the model as it will be

applied to a wheelchair, which is not supposed to be

as fast as a car.

Thereafter, a navigation system will be implemented

using an adaptive Neuro-Fuzzy Inference System to

generate the assistive control reference.

2 DRIVING STATE INFERENCE

USING BAYES NETWORK

This section will brieﬂy present the model imple-

mented for the driving task inference, as the result is

an input of the Neuro-Fuzzy model.

The model considers a scenario where the

wheelchair’s user moves from a point A to a

point B following a straight line, and an obstacle

is standing on the way, or moving across the path

(ﬁgure 1).

Hence the following four states are thus deﬁned:

• State 1 (Straight line driving task): the vehicle

goes from A to B without any obstacle on the path

• State 2 (Slow change of trajectory): the user

avoids an obstacle which is far ahead (or moving

slowly).

• State 3 (Rapid manoeuvre): the user initiates an

emergency collision avoidance manoeuvre.

• State 4 (Recovery, from rapid manoeuvre): the

driver deﬁnes a new trajectory to the destination

after an obstacle avoidance manoeuvre.

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

124

Figure 1: Scenario representation.

The structure of the model is assumed to be known

and is presented on ﬁgure 2.

Figure 2: Bayesian network model.

Therefore, the model estimation is reduced to the

estimation of the parameters. The judgment X(t) is

a discrete variable representing the actual state of the

driving task, while the observation Y(t) is the distri-

bution of observed information, namely the command

signal for the linear velocity and the command signal

for the angular deviation which are both measurable

at the output of the controller (in this case, a joystick).

MURPHY (Murphy, 2007) developed a Bayes Net

Toolbox for Matlab, which implements the EM (Ex-

pected Maximization) algorithm to ﬁnd local optimal

maximum likelihood estimates of the parameters. The

model training is performed as the combination of

time-series judgment states and observed data at that

time.

The results shown in ﬁgures 3, 4, 5 and 6 compare

the data from the controller and the states inferred by

the model. This validates the aptitude of the model at

inferring on human behaviour in wheelchair driving

activity.

3 ANFIS MODEL

3.1 The Theory Reviewed

A simple representation of ANFIS architecture (Jang,

1993) is presented in ﬁgure 7.

(a) Test data

(b) Estimated state

Figure 3: State 1: Normal path following.

(a) Test data

(b) Estimated state

Figure 4: State 2: Slow trajectory change.

The functions of the layers as deﬁned by DENAI

et al. (Denai et al., 2004) are:

• Layer 1 generates the membership grades

• Layer 2 generates the ﬁring strengths by multiply-

ing the incoming signals and outputs the t-norm

operator result.

• Layer 3 normalizes the ﬁring strengths

• Layer 4 calculate rules outputs based on the con-

sequent parameters

• Layer 5 computes the overall output as the sum-

ADAPTIVE COMPENSATION SIGNAL FOR A WHEELCHAIR CONTROL USING ANFIS MODEL

125

(a) Test data

(b) Estimated state

Figure 5: State 3: Rapid manoeuvre.

(a) Test data

(b) Estimated state

Figure 6: State 4: Recovery.

mation of incoming signals.

There are basically two classes of learning algorithm

for ANFIS architecture:

• In the forward pass, outputs of the nodes go for-

ward until the fourth layer where least square

methods are used to identify the conclusions.

• In the backward pass of the hybrid algorithm,

errors signals are now propagated backwards

and the gradient descent method updates the an-

tecedent parameters.

Figure 7: ANFIS architecture.

3.2 Model Deﬁnition

Our model is supposed to generate an adequate con-

trol signal, given the driving state of the driver and the

environmental inﬂuences. The only external (major)

factor considered in this study is the obstacle, which

may be mobile or static; therefore, the inﬂuence of the

surrounding environment is reduced to the distance

between the vehicle and the obstacle. Note that the

considered distance is measured from the vehicle to

the closest impact point of the vehicle, following the

displacement direction of the chair (see ﬁgure 8).

Figure 8: distance measured between the vehicle and the

obstacle.

The general scheme of the model to implement for

this task (ﬁgure 9) is very close to the one proposed

by NEFTI et al. (Nefti et al., 2001). Four possible

responses (each one corresponding to a driving state

of the wheelchair driver) are computed given the dis-

tance to the obstacle, and the driving state inferred by

the Bayesian network will choose the adequate out-

put.

The distance to collision is considered to be the

only input of the different systems, and each system

will output two signals: the linear speed and the an-

gular deviation. Now, lets deﬁne the rules of the dif-

ferent systems.

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

126

Figure 9: ANFIS model general scheme.

State 1. In the ﬁrst state, the driver follows a straight

line to the target. Therefore, the angle is supposed

to remain zero, and the linear velocity is supposed to

remain at an average value depending on the users be-

haviour.

Angle ≈ 0

Linear velocity ≈ V

avg

State 2. The second state corresponds to the slow

change of trajectory. Here, we consider the two out-

puts: the speed and the angular deviation.

If the obstacle is far, then the speed is normal and the

direction changes very slowly

If the obstacle is close, then the speed is slow and the

direction changes slowly

State 3. The obstacle is relatively very close, and per-

haps moving towards the vehicle, the driver initiates

emergency obstacle avoidance.

If the obstacle is far, then the speed is slow and the

direction changes slowly

If the obstacle is close, then the speed is very slow

and the direction changes quickly

State 4. It would not be appropriate to consider the

distance to collision in this case as the vehicle has

overcome the obstacle and the driver tries to deﬁne

the new trajectory to the target. Therefore, the goal

is to get back to the normal driving path, which can

be deﬁned by normal linear velocity and no angular

deviation (just like in state 1).

4 RESULTS

The experiments conducted permitted to gather the in-

formation about the X and Y position of the vehicle,

as well as the deviation angle; those parameters are

used to compute the distance between the vehicle and

an obstacle on its way.

As mentioned in the previous section, only the

second and third phases of the driving behaviour are

Figure 10: Initial membership functions.

subject to a Neuro-Fuzzy controller, as the two other

states will have constant references.

The toolbox generates initial membership func-

tions (ﬁgure 10) and uses the backpropagation algo-

rithm for parameter tuning. The resulting member-

ship functions of ﬁgures 11, 12 and 14 represent the

membership grades as functions of the distance to col-

lision.

Figure 11: Final membership functions: state 2.

In ﬁgure 11, the bell-shaped function representing

the set of close distances has its center at about 2.7m

while the distance is more likely to be considered as

far from the value of about 3.1m.

Even though the ﬁrst function of ﬁgure 12 looks

wider (regarding the input range) than that of the ﬁg-

ure 11, the range is considerably smaller (from 0.7m

to about 1.4m).

The two functions depicted in ﬁgure 14 have al-

most same widths of about 0.25m. From 1.2m to

0.85m, the user considers the distance to be critical;

moreover, the narrowness of the function and the very

small overlap with the function express a decision

making process involving a limited output range.

Figures 13 to 15 compare the data from the user

and the data generated by ANFIS models. The re-

sponse of the ANFIS model in the third state shows a

ADAPTIVE COMPENSATION SIGNAL FOR A WHEELCHAIR CONTROL USING ANFIS MODEL

127

Figure 12: Membership functions for angle inference in

state 3.

(a) Signal from the Joystick

(b) ANFIS generated signal

Figure 13: State 2: Slow trajectory change.

delay in reducing the linear velocity, however the an-

gle reference tries to avoid the obstacle (changing the

direction) as soon as the state is detected. Moreover,

the model initiates a reduction of the speed when the

driving task persists. Finally, we can note the return

to the normal behaviour (maximum speed and no de-

viation) when the vehicle has overcame the obstacle.

Figure 14: Membership functions for speed inference in

state 3.

(a) Signal from the Joystick

(b) ANFIS generated signal

Figure 15: State 3: Collision avoidance manoeuvre.

5 CONCLUSIONS

The results show the possibility of an assisted con-

trol adapted to the behaviour of the user. The imple-

mented ANFIS model is to serve as generator of the

reference signal for the user input. The combination

ICINCO 2011 - 8th International Conference on Informatics in Control, Automation and Robotics

128

of the two signal is the input to the wheelchair sys-

tem. A future improvement of the overall system may

include a state estimator at the output of the system,

in order to implement a feedback controller using op-

timal control theory.

ACKNOWLEDGEMENTS

The authors would like to thank the Tshwane Uni-

versty of Technology (TUT) for the ﬁnancial support,

and also the French South African Institute of Tech-

nology (F’SATI) for providing the necessary tools

permitting to conduct this work.

REFERENCES

Araghi, M. H. and Menhaj, M. B. (2008). A combined col-

lision cone and fuzzy based obstacle avoidance algo-

rithm for the small size soccer robots. International

Conference on Automation, Robotics and Control Sys-

tems, pages 151–158.

Ayari, I. and Chatti, A. (2007). Reactive control using be-

haviour modelling of a mobile robot. International

Journal of Computers, Communications & Control,

Vol. II, pages 217–228.

Bouslimi, W., Kassaagi, M., Lourdeaux, D., and Fuchs, P.

(2005). Augmented nave bayesian network for driver

behaviour modelling. Intelligent Vehicles Symposium,

IEEE Proceedings.

Burnham, G. O., Seo, J., and Bekey, G. A. (1974). Identiﬁ-

cation of human driver models in car following. IEEE

transactions on automatic control.

Chakravarthy, A. and Ghose, D. (1998). Obstacle avoidance

in a dynamic environment: A collision cone approach.

IEEE transactions on systems, Man and cybernetics

part A: systems and human, Vol. 28.

Dario Salvucci, E. R. B. and Liu, A. (2001). Toward an

integrated model of driver behaviour in a cognitive ar-

chitecture. Transportation Research Record.

Denai, M. A., Palis, F., and Zeghbib, A. (2004). Anﬁs based

modelling and control of non-linear systems: a tuto-

rial. IEEE International Conference on Systems, Man

and Cybernetics, pages 3433–3438.

Emam, H. (2010). dynamic modelling and control for the

electronic wheelchair. PhD thesis, University of Ver-

sailles, France.

Fiorini, P. and Shiller, Z. (1998). Motion planning in dy-

namic environment using velocity obstacles. Interna-

tional Journal of Robotics Research, Vol. 17, pages

760–772.

Jang, J.-S. R. (1993). Anﬁs: Adaptive-network-based fuzzy

inference systems. IEEE Trans. on Systems, Man, and

Cybernetics, vol. 23, pages 665–685.

Kim, J.-H., Hayakawa, S., Suzuki, T., Hayashi, K., Okuma,

S., Tsuchida, N., Shimuzu, M., and Kido, S. (2004).

Modelling of drivers collision avoidance behaviour

based on piecewise linear model. IEEE conference

on decision and control.

Koashi, M., Hayakawa, S., Suzuki, T., Okuma, S.,

Tsuchida, N., Shimuzu, M., and Kido, S. (2003).

Measurement and modelling of collision avoidance

behaviour of drivers using three dimensional driving

simulator. SICE annual conference in FUKUI.

Koren, Y. and Boreinstein, J. (1991). Potential ﬁeld meth-

ods and their inherent limitations for mobile robot

navigation. Proceedings of the IEEE Conference on

Robotics and Automation, pages 1398–1404.

Kuge, N., Yamamura, T., Shimoyama, O., and Liu, A.

(1998). A driver behaviour recognition method based

on a driver model framework. society of automotive

engineers.

Langer, D., Rosenblatt, J., and Hebert, M. (1994). A

behavior-based system for off-road navigation. IEEE

Trans. Robotics and Automation, Vol. 10, No. 6, pages

776–782.

Masoud, A. A. (2002). Evasion of multiple, intelligent pur-

suers in a stationary, cluttered environment: A har-

monic potential ﬁeld approach. Proceedings of the

IEEE International Symposium on Intelligent Control.

Murphy, K. (2007). Bayes net toolbox for matlab.

http://code.google.com/p/bnt/.

Nefti, S., Oussalah, M., Djouani, K., And Pontnau, J.

(2001). Intelligent adaptive mobile robot naviga-

tion. Journal of Intelligent and Robotic Systems, pages

311–329.

Rosenblatt, J. K. (1997). DAMN: A Distributed Architec-

ture for Mobile Navigation. PhD thesis, The Robotics

Institute of Carnegie Mellon University, Pittsburgh,

Pennsylvania.

Shimizuhira, W., Fujii, K., and Maeda, Y. (2004). Fuzzy

behavior control for autonomous mobile robot in dy-

namic environment with multiple omnidirectional vi-

sion system. Proceedings of IEEE /RSJ International

Conference on Intelligent Robots and Systems.

Suzuki, T., Yamada, S., Hayakawa, S., Tsuchida, N., Tsuda,

T., and Fujinami, H. (2005). Modelling of drivers

collision avoidance behavior based on hybrid system

model an approach with data clustering.

Tezuka, S., Soma, H., and Tanifuji, K. (2006). A study

of driver behaviour inference model at time of lane

change using bayesian networks. Proceedings of IEE,

ICIT.

Thakur, K. P. (1997). Simulation of driver behaviour as a

function of driver error and driver daydream factor.

Yinka-Banjo, C. O. (2010). Collision avoidance in un-

structured environments for autonomous robots: A

behavioural modelling approach. Master’s thesis,

African Institute for Mathematical sciences, Cape

Town, South Africa.

ADAPTIVE COMPENSATION SIGNAL FOR A WHEELCHAIR CONTROL USING ANFIS MODEL

129