USER ADAPTATION IN A PEDESTRIAN GUIDANCE SYSTEM FOR
THE BLIND
Vivien Guillet
∗,∗∗
, Beatrice Rumpler
∗
, Jean-Marie Pinon
∗
∗
LIRIS-INSA CNRS FRE 2672
INSA - Btiment Blaise Pascal
7, Avenue Jean Capelle
F-69621 Villeurbanne Cedex - France
∗∗
EO-EDPS
69, rue Gorge de Loup
69009 Lyon - France
Keywords:
Adaptation, guidance, user modeling, stereotype, impairement
Abstract:
Among today’s emerging pedestrian guidance systems (i.e. able to automatically answer questions of the
form “How should I go from localization a to localization b ?), some are dedicated to blind people; fewer if
any focus on impaired people as well as people without any deficiency. Contrary to this, the “ouvej” system
relies on representing impairments as a part of the user’s profile in order to be adaptable to any kind of
user. Furthermore, we consider that benefits of the user modeling involved in such adaptation mechanisms
go beyond the only impairment category in order to respect the user’s preferences in itineraries and itineraries
descriptions, by modeling itinerary environment. Whereas existing adaptive system often rely on the use of
explicit feedback from the user, we propose to limit adaptation-related interactions by using implicit user
evaluations inferred from observations of users interactions with the system. This article focuses on blind
pedestrian users using a vocal interface. After detailing our approach, we detail our itinerary and user models;
Then the implicit as well as explicit itinerary evaluation mechanisms are described; Next sections focus on the
actual adaptation and expected benefits of inferences between different users; Our prototype is then described.
We conclude by presenting our first observations, future work and evaluation.
1 INTRODUCTION
Among numerous recent guidance help systems for
the pedestrian research works or products, often in-
spired of existing guidance system for vehicles, some
are dedicated to visually impaired people (“Vikktor
Trekker” by Visuaid; works of the polytechnic in-
stitute of Lausane (Rouller et al., 2002) and LIMSI
(Gaunet and Briffault, 2001)). Such prototypes typi-
cally involve the Global Positioning System technol-
ogy for user localization, exploiting electronic map
for the guidance.
Our study is part of the “Ouvej” project
1
which
goal is to provide a general ubiquitous guidance help
system for any pedestrian. In short, such a system
should be able to automatically answer questions of
the form “How should I go from localization a to lo-
calization b ?”. User’s potential deficiency has been
taken into account since the beginning of this project:
the answer to the prior question can’t be answered the
1
directed by Mission Handicap (Universit Claude
Bernard, Lyon 1), assisted by Liris-Insa (INSA-Lyon) and
E.O.-E.D.P.S, and founded by Region Rhne-Alpes
same way for someone without any deficiency than
for someone moving in a wheelchair. Moreover, the
best medium for providing such information might
not be the same for different kind of user. Thus, em-
phasis has been put on the necessary user adaptation
of the end-user interface as well as the content of the
assistance itself (i.e. description of the itinerary). This
article addresses the particular issue of the visually
impaired user using a vocal (telephonic) interface.
After detailing our approach, we detail our itinerary
and user models; Then the implicit as well as explicit
itinerary evaluation mechanisms are described; Next
sections focus on the actual adaptation and expected
benefits of inferences between different users; Our
prototype is then described. We conclude by present-
ing our first observations, future work and evaluation.
2 APPROACH
Interviews and observations have shown that blind
people are often reluctant to use a phone or any talk-
ing device while walking and consider it dangerous.
69
Guillet V., Rumpler B. and Pinon J. (2004).
USER ADAPTATION IN A PEDESTRIAN GUIDANCE SYSTEM FOR THE BLIND.
In Proceedings of the Sixth International Conference on Enterprise Information Systems, pages 69-74
DOI: 10.5220/0002650300690074
Copyright
c
 SciTePress

The main reason given for that is that the cognitive
load is generally too heavy to handle both activi-
ties. Moreover, it is often impossible to use a phone
while using a cane or following a guide dog. Thus,
the moves using a voice-based interface for the blind
must be achieved in a set of two steps : (1) the sys-
tem is asked for an itinerary and replies; (2) The user
follows the given itinerary description.
Thus, the main concern about our user adaptation
mechanism is to provide a description of an itinerary
being (1) as detailed as possible in order for the user
to be able to follow it, and in the same time (2) as
short as possible in order to be easily memorized by
the user. This optimal detail level for a description is
related to :
• Each user (different users need different level de-
tails);
• Each itinerary, which complexity isn’t a priori
known.
In order to adapt the itinerary description to user’s
needs, the user could be asked for desired detail level
by the system. But then, it should be asked for each
itinerary, which is not a suitable solution : the benefits
of adaptation would be cancelled by the extra time
involved in this interaction. Contrary to this approach,
we consider that the user adaptation can be achieved
using a minimal amount of user explicit feedback: as
we are fully aware of the whole history of the user’s
interactions with the system (“use traces”) should be
sufficient to set optimal parameters for each user and
each itinerary detail level.
We propose to use this knowledge to determine the
best parameters for an user according to its past use
traces, and to take part of other user’s traces in or-
der to predict those parameters for the new user of
the system, and for each itinerary. Such user mod-
eling related techniques have proven to be useful for
a wide range of applications, performing tasks that
can be classified into document filtering, document
personalization and interface adaptation (Kobsa and
Wahlster, 1989) (Leake, 1996).
3 VOCAL USER INTERFACE
DESCRIPTION
The use traces study is closely related to the actual
user interface. However, the implicit feedback pro-
vided by the user while interacting with the system
not necessarily implies the use of a single interface.
Indeed, the same simple feedback is expected to be
collected with other user interfaces of the system, as
those interfaces have to be designed to do so.
Once asked for an itinerary, the system replies by
Figure 1: Simplified vocal interface flow
reading a sequence of itinerary segment descriptions
2
or steps. The simplified vocal user interface (limited
to itinerary description), is described figure 1. If the
user doesn’t interact at this point, the normal flow is
used. At any time, the user can ask for the next step,
the previous step, or a step specified by its number.
Use traces consist in a record of these actions,
labelled with the ongoing step number. We con-
sider that the analysis of the user circulation into the
itinerary description represents its implicit evaluation:
a poorly formulated description will not either not be
well remembered by the user, or lead to difficulties
in the moving process, leading the user to ask for the
segment description again. On the contrary, a well
formulated itinerary description will rarely be asked
many times, and often passed. Use traces exploitation
is detailed in section 5.
However, the interface permits the use of explicit
user feedback for critical adaptation tasks. Such a
feedback is expected to be used in the case of an
impossible itinerary: for instance, blind people often
won’t accept to cross a road that is not “securized”
(e.g. that might have traffic light, but no associated
vocal message)
3
; thus, this crossing will be consid-
ered impossible for an user, regardless of the partic-
ular itinerary in which it takes place. Such an adap-
2
This itinerary decomposition into segments results
from the itinerary computation (see 4.1)
3
although some might
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
70

tation could be handled using the same mechanism
as above (use implicit feedback contained in the use
traces), but could result in dangerous situations if the
user is unaware of the possible adaptation.
4 USER AND ITINERARY
MODELING
4.1 Itinerary model
The modeling of an itinerary has been achieved by
studying real-life itinerary descriptions dedicated to
french-speaking blind people
4
. These descriptions
have been reformulated using a set of generation rules
of the form defining an itinerary description gram-
mar, which use a subset of change of localization (i.e.
movement) classes (Muller and Sarda, 1998) (Ma-
thet, 1998), as well as localization categories (Sarda,
1997).
5
1
2
3
4
Figure 2: Minimal movement types set
Each step of an itinerary can be fully defined by an
action and a final localization, both being potentially
implicit :
• An action is defined by associating an item (a
road, a pedestrian crossing, a sidewalk, a stair, a
doors, etc.) optionally labelled with item categories
(buildings, stairs, etc.) with a movement: go along-
side; go across; follow; go in a direction (see fig-
ure 2). Both action and movement being optionally
given a set of descriptions.
4
Thanks to E.O.-E.D.P.S.
5
The correctness of our modeling could be closely re-
lated to the language used in the corpus.
• A final localization is an item, and a set of optional
descriptions.
As we mentioned above (section 2), a fully detailed
description will be more precise, being more helpful
for the user to conceive an inner personal map of its
environment while making it more difficult for him to
remember the described itinerary. However, the short-
est description might not be the best one, as it could
lead to misconceptions, even at the scale of a single
itinerary.
Thus, the output description can be adaptated to
the user’s need by including or not (1) optional de-
scriptions of an item that may include relative local-
ization of the item according to the user’s position
(“the crossing situated at your left”), localization of
the item according to other items (e.g. “the stairs sit-
uated in front of the tramway rails”); (2) reference to
user awareness of an item, based on a prior use found
in user’s history; (3) free-form optional description.
The adaptation process includes also the choice of the
form of the sentence used for an action description
(for instance: explicitation of optional actions or lo-
calization).
The generation process of the description of an
itinerary uses a set of parameters which are neces-
sary directly mapped on each of these options (e.g.
maximal length of the description text, description of
specific items); Hence, the referred as (description)
construction parameters.
Due to the nature of the description corpus, our
itinerary modeling is closely related to blind people’s
moves. Sighted people, for instance, might find it pe-
culiar to be told to walk until the lowering of a side-
walk (which is especially useful when followed with
a cane) as it is not the kind of item they use in their
everyday moves. Such users will prefer using visual
items, use building names, etc.
However, the above itinerary modeling can be eas-
ily extended to non-blind people: instead of conceiv-
ing specific itineraries for each user categories (that
are hardly to be defined), an adaptation process for
the itinerary building can be achieved in filtering the
items and actions according to the user’s actual mov-
ing possibilities provided be explicit feedback
6
. In-
deed, the limit is thin between sighted and visually
impaired people as visual impairment itself not only
covers a wide range in intensity, but also in categories.
Furthermore, given a visual deficiency, some people
might consider themselves blind or not. Thus, we
consider possible for two blind people to have less in
common than a sighted and a blind people.
As a result, a step is a bag of (action,item) set (and
their associated descriptions). Given two different
users, the same item is possibly optionally used for
6
However, some actions forms have to be added to
widen the range of possible itinerary descriptions
USER ADAPTATION IN A PEDESTRIAN GUIDANCE SYSTEM FOR THE BLIND
71

the first and mandatory for the other, as well used, or
not, in an action.
4.2 User Model
As suggested by their name, user modeling based sys-
tems often introduces attributes in the user profile that
are not closely related to way the user interacts with
the system, especially if impaired people were ex-
pected to be involved (Jeribi, 2001). Such attributes
might include user’s age, sex, impairment specifica-
tion, etc. This knowledge permits direct and easy
adaptation, and has proven to be useful (Kobsa, 2001).
However, the use of such a knowledge, being soci-
ological, cognitive or physiological results in supple-
mentary constraints in the reasoning that blur analy-
sis of the system reasoning capabilities. For instance,
observation of visually impaired people’s move have
suggested user’s age as a good factor for determin-
ing the detail level of an itinerary description: the
youngest need less optional descriptions and memo-
rizes more easily a given description than the oldest.
But a closer look shows that this difference might in
fact be related to the locomotion school he did belong
(or not), this criterion being not necessarily related to
the person’s age, but merely to the age the person be-
came blind.
On the contrary, considering the only use of the
system by the user has the benefits of (1) helping to
avoid the temptation of adding a priori and uncertain
knowledge about the user; (2) making it easier to ob-
serve the dynamical functioning of the reasoning sys-
tem itself (Pierre-Antoine Champin, 2002).
5 ITINERARY STEPS
EVALUATION BY USERS
User adaptation of step and step description relies on
user’s evaluation of itineraries steps. These evalua-
tions are constructed whether explicitly or implicitly
(i.e. built from the user traces), as shown in 3; Explicit
evaluation are used for critical tasks (the user reports
an itinerary he isn’t capable to follow), implicit eval-
uation for evaluating the descriptions. Evaluations of
itinerary as well as descriptions is done as a whole,
subsequent reasoning being used to determine its in-
ner element contribution in the evaluation.
Given an itinerary search session and the corre-
sponding itinerary formed of the ordinated step de-
scriptions of steps, the session use trace is a time-
stamped sequence of pairs of steps and step descrip-
tions actually heard by the user.
This trace is interpreted as follows : starting from
the second occurrence of a step, the associated eval-
uation will be removed one point. To the contrary,
description of steps that are heard only once will gain
a point.
As the description form of a given description is
determined by the values of its construction param-
eters {p
1
. . . p
n
} such as optional items and actions
descriptions, optional item localization, free-form op-
tional description and form of the sentence (described
in 4.1), the role they play in the positive as well as
negative evaluation of a step description has to be de-
termined.
Thus, as possible values of each construction pa-
rameter p
i
consists in a set of m discrete values
p
i,1...m
, evaluations are directly used for the evalu-
ation of each construction parameter value. Evalua-
tion for parameter value p
i,j
is noted E(p
i,j
). Each
evaluation given by an use trace is added to previous
evaluations of each parameter-parameter value pair,
as illustrated figures 3 and 4. A confidence level for
each parameter value is given by the number of eval-
uations.
trace param. value evaluation
t
1
v
p
i
,1
-2
t
2
v
p
i
,2
-2
t
3
v
p
i
,1
1
t
4
v
p
i
,3
-3
t
5
v
p
i
,3
1
t
6
v
p
i
,3
1
Figure 3: Example: evaluations for different parameter val-
ues v
p
i
,1...3
, for p
i
trace E(p
i,1
) E(p
i,2
) E(p
i,3
)
t
1
-2
t
2
-2
t
3
1
t
4
-3
t
5
1
t
6
1
total -1 -2 -1
evaluations nb 2 1 3
Figure 4: Example : addition of step evaluations on related
parameter values p
i,1,...,3
, for a given parameter p
i
The eventuality must be considered of a step de-
scription being intrinsically difficult to follow, either
due to an inaccurate transposition of the real environ-
ment into its representation in the itinerary model (in-
cluding possibility of changes in topology since this
transposition), or to actual difficulty of the move. In
order for this difficulty not to be attributed to the con-
struction parameters in the case of a negative evalua-
tion of the step, the evaluation must be associated not
only with a set of construction parameters, but also
with the step whose description is evaluated. Hence,
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
72

the cumulated evaluation serves as a weight for con-
struction attributes evaluation.
6 ADAPTATION TO THE USER
The direct feedback provided by the user (explicit
evaluation) in the case of a an impossible itinerary is
stored as a (action, item, truth value), the truth value
being a boolean value representing the user’s actual
possibility of a given action on a given item. Thus,
corresponding adaptation consists in a simple filter-
ing on the item and actions defining an itinerary step,
ensuring a step to be suitable for an user.
Adaptation of the itinerary description is basically
used by using the most adapted construction parame-
ters values while building the itinerary description for
a given user. These values are selected using the im-
plicit evaluation process described above. Note that
the implicit nature of user evaluations not necessarily
implies the indirect parameter evaluation described
above; however, both help achieving the same goal of
adapting to the user with a minimum amount of feed-
back. For this reason, mechanism directly involved in
the adaptation filtering does not dependent on specific
evaluation mechanism.
Construction parameters includes low-level param-
eters (directly mapped to the elements of the itinerary
model), as well as higher level parameters, such as :
• Maximal length of the description text;
• List of reliable item categories for optional descrip-
tion;
• Inclusion of free form optional descriptions;
• Inclusion of references to user’s past itineraries, for
each item known by the system to have been used
by the user (as stored in use traces).
As the system relies on submitting various parame-
ter values to each user, it is possible that invariable in
time use of the highest-ranked construction parame-
ters would result in keeping the first good value. Thus,
special care must be taken in the choice of the con-
struction parameters value in order to provide an eval-
uation of each one, while keeping in mind that the
benefits of learning from an user will be lost to him
if the system keeps using alternative construction pa-
rameters values.
7 FURTHER INFERENCES
As a common difficulty of user adaptation resides
in the model instance initialization phase (often re-
ferred as “cold start”), subsequent reasoning must be
involved. As our system is to be used by many users,
inferences can be drawn between users.
The stereotype approach (Rich, 1979) reposes on a
classification of users according to their profiles (user
model instances). As a result, a default profile is
established for each class, using a-priori knowledge
(static stereotypes) or clustering algorithms. Hence, a
new user is associated a class, thus an associated de-
fault profile. However, stereotypal reasoning is not
limited to user’s profile initialization: many infer-
ences can be draw between domain objects and their
attributes (e.g. an item; its label) and user attributes
(e.g. user evaluations of construction attributes. How-
ever, such inferences are beyond the scope of this ar-
ticle (see (Kay, 1994) for further explanations).
8 PROTOTYPE
The prototype is build upon a web server, a geo-
graphical database represented in the form defined
by the itinerary modeling, and an user model (UM)
database, containing evaluations of construction pa-
rameters, use traces, and the list of possible (item, ac-
tion) pairs.
The adaptive mechanisms of the system (user mod-
eling relative part of the prototype represented fig. 5)
are currently not implemented, but the vocal interface,
as well as localization of users using the GPS tech-
nology are functional. Thus, the corpus used in our
itinerary and environment modeling has been already
tested on blind people inside Campus de la Doua
(Lyon 1) , as well as the overall functioning of the
prototype.
Both interface and content (itinerary and itinerary
description) can be described using XML grammars,
thus permitting to realize a two pass adaptation based
on w3c standards (Heckmann and Kr
¨
uger, 2003), al-
though only the content part has been discussed here.
When asked for an itinerary by an user, the systems
extracts a sequential set of steps from the geograph-
ical database presented as an itinerary description
XML document(1). An itinerary description transfor-
mation style sheet (stored in XSLT) is extracted from
the UM database, which is applied to (1);
As the system is not limited to its use with the vocal
interface we described but have to be multi-modal, an-
other transformation step is done by applying an out-
put format style sheet to the form-independent result
of the previous transformation in order to generate the
appropriate format : VoiceXML for a vocal interface,
HTML for desktop computers (small sized for PDAs),
etc.
USER ADAPTATION IN A PEDESTRIAN GUIDANCE SYSTEM FOR THE BLIND
73

Figure 5: Simplified interface and document adaptation
framework
9 CONCLUSIONS
The adaptation process described in this paper will be
studied besides its actual implementation in order to
determine its appropriate work conditions. Dynami-
cal aspects of this study will include analysis of the
number of generation parameters and the number of
generation parameters values relatively to the conver-
gence of a profile; sensibility to initial conditions will
be observed in order to predict the role of stereotypes
as a possible profile initialization. These observations
will be used as a starting point for potential inter-user
reasoning.
Preliminary tests using our prototype have shown
the accuracy of the corpus used for itinerary model-
ing; the implementation of the user modeling in the
prototype will permit evaluation of this modeling it-
self, as well as our user-evaluation process for the par-
ticular guidance task. Clustering techniques applied
on the resulting user profile corpus will be used in
order to compare preexisting classes with system-use
related class, and evaluate their accuracy.
REFERENCES
Gaunet, F. and Briffault, X. (2001). Mthodologie de spci-
fication pour un systme informatique de guidage ver-
bal : le dplacement des dficients visuels en ville. In
de Psychologie, S. F., editor, EPIQUE 2001, Journes
d’Etudes en Psychologie Ergonomique, pages 133–
143.
Heckmann, D. and Kr
¨
uger, A. (2003). A user modeling
markup language (userml) for ubiquitous computing.
In Brusilovsky, P., Corbett, A. T., and de Rosis, F.,
editors, User Modeling 2003, 9th International Con-
ference, UM 2003, pages 393–397. Springer.
Jeribi, L. (2001). Thse. PhD thesis, Insa-Lyon.
Kay, J. (1994). Lies, damn lies, and stereotypes: pragmatic
approximations of users.
Kobsa, A. (2001). Generic user modeling systems.
Kobsa, A. and Wahlster, W. (1989). User Models in Dialog
Systems. Springer Verlag, Heidelberg, Berlin.
Leake, D. B. (1996). Cbr in context: the present and the
future, in case based reasoning experiences, lessons
and future directions. In Park, A. P. M., editor, aa,
pages 3–30.
Mathet, Y. (1998). Un formalisme pour le dplacement. In
actes de TALN ’98.
Muller, P. and Sarda, L. (1998). Reprsentation de la sman-
tique des verbes de dplacement transitifs du franais.
In Actes de la Confrence Traitement Automatique du
Langage Naturel (TALN’98), Paris, Juin 1998, pages
127 – 147.
Pierre-Antoine Champin, Y. P. (2002). Mod
´
eliser
l’utilisateur ou l’utilisation. In Actes de DVP 2002.
Rich, E. (1979). User modeling via stereotypes. Cognitive
Science, 3, pages 355–66.
Rouller, J., Perrottet, D., Ladetto, Q., and Merminod, B.
(2002). Faciliter le dplacement des aveugles avec une
carte numrique et une interface vocale. In Geoinfor-
mationseminar, Positionsabgleich mit der Karte, Map
Matching Techniken.
Sarda, L. (1997). La smantique des verbes de dplacement
transitifs directs. tentative de description du proces-
sus de localisation. In Actes de la Sixime cole d’t de
l’Association pour la recherche cognitive.
ICEIS 2004 - HUMAN-COMPUTER INTERACTION
74