Behavior-Based Indoor Navigation - Ambient Intelligence for Scientific Discovery - page 270

Fig. 2. Original and segmented images.

simple thresholding is enough to appropriately segment the green areas of the
image. Once the image is segmented, we need a method to classify an image as
containing, or not, a panel; instead of using image correlation functions we ap-
plied a Multi Layer Perceptron (MLP) neural network trained using as an input
vector the quadratic weighted sums of 20

20 sized image blocks, resulting in 28
input neurons [43]. The training image set contained the inputs of 320 images,
taken from the nominal trajectory maintaining constant pan, tilt and zoom val-
ues. The neural net has a single hidden layer of 3 neurons and the output is a
binary vector that classifies the image as containing, or not, a panel. Applying a
leaving-one-out (LOO) validation technique, and after a training period of 1000
epochs for each set, we obtained a validated performance of 96 . 25%. The MLP
is improved by looking for appropriate input weights using a Genetic Algorithm
[32] and associating to the individuals as a fitness function the LOO performance
of the MLP itself, raising the accuracy up to 97 . 8%.
The emergency exit panel recognition module, based on the MLP, gives, as
an output, the mean value of the last 10 images. This value is a measure of the
confidence level ( cl ) of the recognition process. Therefore, it will send an output
“1” indicating that a new landmark has been detected only after 10 positive
identifications. When this occurs, the proper actions would be taken.
Identification of emergency exit panels is more dicult than of corridors
due to their lack of duration in time. In spite of the defined confidence level,
the positive identification relies on a few snapshots. To cope better with this
problem, we made the confidence level affect the global translational velocity of
the robot according to the following expression: v =(1

×

v . The aim of this
velocity reduction is to slow down the robot when a panel is being recognized so
that it maintains the panel in range of vision.

− cl ) ∗

Wooden Doors Identification. Doors imply interesting locations in the envi-
ronment. Anyone whose job is to dispatch the daily post, must be able to some-
how identify the different oces that are accessed through doors. The doors in
the robot's environment are made of wood. The objective of this module is to
extract wooden door areas in images taken by the robot and to decide if the
robot is, or not, in front of a door. To do so, the image must be first segmented
by selecting the pixels belonging to a door.
The segmentation method should cope with noisy images; dynamic variations
in the lighting conditions, due especially to the sunlight incoming through the