Behavior-Based Indoor Navigation - Ambient Intelligence for Scientific Discovery - page 272

From this reduced image a number between 0 and 1 measuring the confidence
level when in front of a door is obtained computing the percentage of pixels
labeled as “wooden door” across each column, and calculating the weighted
mean across all the columns, according to the following formula:

i =1 w i p i

i =1 w i

cl wood = (2)

p i being the percentage of pixels labeled as “wooden door” in the i -th column
and w i the weight of the i -th column in the weighted mean. We implemented
a schema in which there are three zones inside the middle columns, according
to the distance from the center of the image. The weight of the columns ( w i )
decreases according to the distance from the middle of the image. Columns near
the outside borders have a weight one third of the columns in the inner zone
and one half of the columns in the intermediate zone. When the confidence level
obtained from an image raises over 0.5, a door has entered in the visual field of
the robot. This number increases over 0.9 when the door fills the center of the
image, and decreases smoothly when the robot surpasses the door. When the
confidence level falls to 0.05 during a predefined number of consecutive images
(to prevent noise effects), the robot has completely passed the door and the door
counter is incremented.

3.3 Active Head Control System
Visual processes are directly related with the task to be performed by the ob-
server. Animate or purposive vision mandates that perception is linked to action
and that vision must be based on expectations [9, 4]. The knowledge on the
environment may somehow condition the interpretation of what is being seen.
Emergency exit panels are small in size and the robot may miss them if the
camera is not adequately focused to the area where the panel is expected to be.
Zoom and tilt values' adaptation, according to distance to the wall, could greatly
improve the robot's behavior when it is looking for these landmarks. With such
active process it is possible to detect emergency exit panels even if the robot
turns aside its nominal trajectory. Note that the tilt adjustment is not desirable
when the robot needs to find doors; when this behavior is required, it is prefer-
able to fix the tilt angle so that door areas fill the upper half of the image. The
wood color area recognition process may also take advantage of the reduced size
of the robot if the tilt position of the camera points straight or slightly down to
the floor; the lower areas of the walls are less sensitive to lighting conditions. All
this emphasizes the principle that vision does not have to function in isolation, if
notaspartofasystemthatinteractswith its environment in order to carry out
its task. The following section describes the selected functions to dynamically
modify the tilt angle and the zoom value.

Distance Based Tilt and Zoom Adjustment. To establish the relationship
between the camera's tilt angle and the distance to the wall, measured by means
of the robot's sonar sensors, instead of imposing a model to the system, we let the