Performance of A Head-Movement Interface For Wheelchair Control

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 4

Proceedings of the 25" Annual International Conference of the IEEE EMBS

Caneun, Mexico * September 17-21,2001

Performance of a Head-Movement Interface for Wheelchair Control


P. B. Taylor', H. T. Nguyen'
'Faculy of Engineering, University of Technology, Sydney, NSW, Australia

Abslracl- Head movement has been used as a control Barea [4] presented initial work on an interface based on
interface for people with motor impairments in a range of electroculography (EOG), where the user gives commands
applications. Chin operated joysticks and switch arrays have to the control system by eye movement. The measured EOG
been incorporated in control systems for electric wheelchairs potentials are passed into an ANN, and the output of this
but have several disadvantages, including being dilfcult to network further analyzed to identify commands. The
operate and aesthetically unattractive. A prototype wheelchair
performance of the ANN was not reported, but results
control interface has been developed that makes use of an
artificial neural network (ANN) to recognize commands given indicating that an able-bodied user used the interface to
by head movement. This paper presents the results of an control a wheelchair along a given path. A similar system,
experimental investigation of the ANN'S performance in terms combining EOG and ekctromyography (EMG) was also
of classification accuracy and delay. It goes on to compare the presented by Law [5], along with results indicating that a
results of disabled with able-bodied users, and assesses the user was able to use the system, to perform 3 tasks.
effect of providing real-time feedback to the user. Min [6] described two wheelchair control interfaces,
The results obtained indicate that ANN techniques can be using head and shoulder movements. One interface allowed
used to classify head movements sufficiently quickly and users to give commands by either tilting or pitching their
accurately to be used in a practical interface. The provision of
head. The other used vertical movements of the shoulders.
graphical real-time feedback does not appear to be crucial, but
may be of benefit for particular cases. Sensor data were translat,ed into commands using thresholds
and heuristics such as hysteresis. Results presented showed
that a group of six disabled users were able to use the
Keywords- head movement, neural network, prototype interfaces to complete a short navigation task. The
classification performance, power wheelchair control. head movement interface was reported to have a recognition
rate of approximately 75%, and the shoulder movement
approximately 70%.
1. INTRODUCTION A prototype wheelchair control interface reported in
Taylor [7] makes use of an ANN to recognize commands
Head movement has been used as a control interface for given by head movement, similar to the system proposed in
people with motor impairments in a range of applications. [Z]. Users give commands to the system by nodding or
Chin operated joysticks and switch arrays have been tilting their head. The movement of the user's head is
incorporated in control systems for electric wheelchairs hut detected by analyzing data from a two-axis accelerometer,
have several disadvantages, including being difficult or collected with a sampling. period of 100 ms. The input to the
unintuitive to operate and aesthetically unattractive. ANN is comprised of a window of 20 samples from each
More advanced interfaces have been developed in axis. When the neural network has recognized a movement,
several projects, but none have yet become commonplace. a simple set of heuristics is used to determine the resulting
Roy [I] showed that Artificial Neural Networks (ANNs) can speed and change of direction of the wheelchair. Results
be used to recognize gestures. This work was able to reach a presented in [7] indicated that ANN techniques could
classification accuracy of 80%. achieve classification accuracies greater than 95% for able-
Joseph [2] described a prototype wheelchair interface bodied users in a laborata'ry setting.
using a triaxial accelerometer to measure head position and This paper presents the results of an experimental
an Artificial Neural Network (ANN) to interpret commands investigation of the performance of a pretrained ANN in
given by head movements. The ANN in this system was terms of classification a.ccuracy and delay. It goes on to
reported to have an accuracy of approximately 92%. compare the results of disabled with able-bodied users, and
Another head movement based interface was described assesses the effect of providing real-time feedback to the
by Bergasa [3]. This system generated commands by using user.
fuzzy logic to recognize the user winking , hiding their lips, Feedback to the user has several forms. Real-time
or tilting or turning their head, based on images captured by graphical displays of the accelerometer data allow the user
a CCD camera. A practical issue reported for this system to track the deviation of their head from the neutral position
was that it performed well in indoor environments with over the preceding 100 samples. Boolean outputs from the
suitable illumination, but this was reduced as light classifier and the numerical output values from the ANN
conditions deteriorated. inform the user of how the classifier is interpreting their
movements.

0-7803-7789-3/03/$17.0002003 IEEE 1590


The performance of the neural network is of data as corresponding to four types of head movement
considerable importance in the development of the system. commands: left, right, forwards or backwards. If the window
The sensitivity and specificity of the classifier and the delay was not classified as belonging to one of the four classes, it
between user's movement and its classification are of was interpreted as being neutral, or equivalently that no
particular significance, as these will affect the safety and command had been given in that window.
utility of the system. Data for each person was collected in two periods of ten
minutes, with the user being prompted to give a specified
movement every 6 seconds. Each specified movement was
11. METHODOLOGY chosen randomly from the following: neutral, forward nod,
forward hold, backward nod, backward hold, left nod, left
Data was collected from six adults, aged between 19 hold, right nod and right hold. The difference between a nod
and 56, with approval from the UTS Human Research and a hold movement is that a nod moves from the neutral
Ethics Committee and informed consent from the position to a maximum value and immediately begins
volunteers. Of these, two had high-level spinal cord injuries returning to the neutral, whereas a hold movement pauses
(C4 and C5) and were not able to use a standard joystick to briefly at the maximum value before returning. A graphical
control a wheelchair. The remaining four did not have real-time display of sensor data and classifier output was
conditions affecting their movement. displayed to the subject for the first period of ten minutes.
A pre-trained ANN was used to classify the windowed This was then hidden for the second period and no feedback
TABLE
I
CONFUSION MATRICES FOR ABLE-BODIED SUBJECTS, WITH AND WITHOUT THE DISPLAY OF REAL-TIME, GRAPHICAL BEDBACK.

-
Actual classification With feedback Actual classification - Without feedback
Able-bodied subjects
Forward Back Left Right Neutral Forward Back Left Right Neutral
Forward 80 0 0 4 6 86 0 0 I 2
Back 0 89 I 0 0 ' 0 94 I 0 0
Expected
classification Left 0
0 85 0 0 0 0 91 0 0
Right 0 0 0 86 I 0 0 0 91 0
Neutral 0 0 0 0 45 0 0 0 0 48

TABLE2
CONFUSION MATNCES FOR DISABLED SUBJECTS, WITH AND WITHOUT THE DISPLAY OF REAL-TIME, GRAPHICAL FEEDBACK.

Actual classification -With feedback Actual classification - Without feedback


Disabled subiccts
Forward Back Left Rieht Neutral Forward Back Left Rieht Neutral
~ ~ ~ ~~ ~ ~

Forward 37 0 0 2 3 41 0 0 0 3
Back 0 42 0 0 0 0 44 0 0 0
Expected
classification 0 0 45 0 0 0 2 38 I 2
Right 2 0 0 35 5 0 1 I 29 II
Neutral 0 0
--
0 0 23 0 0 0
-
0 22

TABLE3
CONFUSION MATRICES FOR ALL ABLE-BODIED SUBJECTS. AND ALL DISABLED SUBJECTS.

Actual classification -Able-bodied Actual classification - Disabled


All results
Forward Back Left Right Neutral Forward Back Left Rieht Neutral
Forward 166 0 0 5 8 78 0 0 2 6
Back 0 183 2 0 0 0 86 0 0 0
Expected
classification Left 0 0 176 0 0 0 2 83 I 2
Right 0 0 0 177 1 2 I 1 64 16
Neutral 0 0 0 0 93 0 0 0 0 45

1591
TABLE
4
S l . U S l l l\'llY, SI'ECIFICITY. POSITIVL PP.EOIClI\'E V A L L S (PPV) A S 0 H I G A ' I IVE PREDICTIVE VAL.CI' (XPV) OFTHI( ARTlblCIAI. NEURAL
N L T H O F X AVF.RA(IED FOR T l l l i 4 POSSlRLt CLASSIFICATIONS.

Sensitivity Specificity PPV NPV

Able-bodied, with GUI 0.967 0.996 0.986 0.990


Able-bodied, without GUI 0.989 0.998 0.995 0.997
Disabled, with GUI 0.929 0.993 0.974 0.981
Disabled, without GUI 0.877 0.992 0.969 0.967
All able-bodied 0.978 0.997 0.990 0.994
All disabled 0.903 0.993 0.971 0.974

TABLE5
MEAN AND STANDARD DEVIATION OF THE DELAY (mr)BETWEEN MOVEMENT ANI1 CLASSIFICATION

Fonvard Back Lett Right

Mean SD Mean SD Mean SD Mean SD

Ablc-bodied, with GUI 1308 149 1278 98 996 100 983 165
Able-bodied, without GUI 1296 86 1212 75 969 IO9 948 94
Disabled, with GUI 1387 753 1624 802 1119 199 . 1178 461
Disabled, without GUI 1378 383 1285 206 1192 822 1260 663
All able-bodied 1302 120 I244 93 982 Ios 965 135
All disabled 1382 589 1445 592 1153 576 1215 557

was provided. IV. DISCUSSION


The raw sensor and classifier data was analyzed then to
produce the results described below. Each IO-minute period Although the number of subjects involved in this trial is
was subdivided into the six-second windows corresponding small, the results obtained lead to several useful
to each specified movement. Windows that contained observations. The performance of the classifier and
obvious errors by the subject were removed fiom the similarities between the ;six subjects appear to validate the
analysis. The start of the movement within each window approach used in developing of the prototype. The results
was determined to be the point where the deviation fiom the also suggest points needing further examination and
neutral position reached 25% of the maximum value on the development.
relevant axis. The classification of the movement was No significant difference was found between specificity
determined as the first classification made by the ANN after of the ANN for able-bodied and disabled subjects. The
the start of the movement. The delays between the start of specificity was high for all of the subjects, indicating that
the movement and the time of classification were then the neural network is unlikely to recognize one type of
calculated, and the confusion matrix relating the expected movement as another.
and actual classification updated. Comparing the delays between the able bodied and
disabled subjects, it appe:ars that although the delay for the
disabled group is generally slightly longer, the difference
111. RESULTS between the two groups is not statistically significant. It is,
however, notable that the able-bodied group had a smaller
The confusion matrices show that the classifier tends to variance in the delay. The implications of this larger
fail to recognize movements at all rather than classify them variance are not clear, but an increase in sample size may
as another type of movement. This is also reflected in the assist in determining this.
high positive predictive value and specificity. The provision of graphical feedback made no significant
As can be seen in the tables, the sensitivity, specificity, difference to the neural network statistics and classification
positive predictive value (PPV) and negative predictive delays for the able-bodied subjects, and only sensitivity and
value (NPV) of the ANN ranged from 0.877 to 0.989, 0.992 NPV were significantly reduced for the disabled subjects.
to 0.998,0.969 to 0.995 and 0.967 to 0.997 respectively, and The majority of the misalassifications leading to this result
mean classification delays ranged from 948 to 1624 ms. were associated with the "Right" class of one subject. This

1592
suggests that real-time feedback is useful where the electrooculography," IEEE Trms. Neural S w e m and
movement being recognized is close to the user's Rehobilitotion Engineering, vol. IO, pp. 209-218,2002.
[SI C. K. H. Law, M. Y. Y. Leung, Y. Xu, and S. K. TSO,"A cap as
comfortable range in that direction. interface for wheelchair control," Proc. IEEENSJ lnternmionol
There is a noticeable difference in the sensitivity of the Con/ on Intelligent Robols o n d s y s l e m , vol. 2, pp. 1439-1444,
neural network between the able-bodied and disabled 2002.
subjects. Again, this is largely a result of the relatively large [6] JLW. Min. K. Lee, S.-C. Lim, and D.-S. Kwon, "Human-
friendly interfaces of wheelchair robotic system for handicapped
number of misclassifications of the Right movement for one persons," Proc. IEEE/RSJ Intonational Con/: on Intelligent
subject, both with and without feedback. This highlights the Robots ondSysrem, vol. 2, pp. 2505-1510,200Z.
need to test the control system with individual potential 171 P. Taylor, H. Nguyen, and A. Craig, "Head Movement
users of the control system. It also suggests that Recognition for Power Wheelchair control," Proc. Annu. Con/
Engineering andphysical Sciences in Medicine, vol. I . pp. 135,
performance of the classifier could be improved by adapting 2002.
it to each particular user.
A practical system should include a fast acting input to
allow the user to react to sudden, unexpected events, for
example, to quickly stop the wheelchair after a
misclassification or in an emergency. The high PPV
obtained indicate that the ANN is able to operate without
producing large numbers of false positives. However, it is
also clear that a small number of false positives can be
expected, and the control interface must be designed to
allow for this. The delays observed indicate that use of the
existing ANN alone is not sufficient to achieve this.
However, use of adaptive algorithms or additional interface
techniques may solve this.

V. CONCLUSION

The results obtained indicate that ANN techniques can


be used to classify head movements sufficiently quickly and
accurately to be used in a practical interface. The provision
of real-time feedback does not appear to be crucial, but may
be of benefit for particular cases.

ACKNOWLEDGMENT

The authors thank Adrian Byak, Allan McCabe and


Matthew Peters for their assistance in the collection of data
for this experiment.

REFERENCES
D.M. P. Roy, M., "Computer recognition of athetoid cerebral
palsy movement using ncural networks," Proc. 16th Con/ IEEE
Engineering in Medicine and Biology Society, vol. I, pp. 458-
459,1994.
T. Joscph and H. Nguycn, "Neural network control of
wheelchain using telemcvic head movement," Pmc. 20th Con/:
IEEE Engineering in Medicine and Biology Society, vol. 5 , pp.
2731-2733, 1998.
L. M. Bcrgasa, M. Mazo, A. Gardel, R. Barca, and L. Boquete,
"Commands generation by facc movements applied LO the
guidance of a wheelchair for handicapped people," Proc. 1Slh
CO$ Panern Recognition, vol. 4, pp. 660-663,2000,
R. Barea, L. Boquctc, M. M a o , and E. Lopez, "System for
assisted mobility using eye movementS based on

1593

You might also like