Online Feature Weighting for Human Discrimination in a Person Following Robot - Foundations on Natural and Artificial Computation - page 227

3 Experimental Results

We have tested our proposal over two video sequences captured while the
robot is performing the person-following behaviour and when there are
several distractors trying to make the system fail. These videos were logged in
the Department of Electronics & Computer Science, at the University of
Santiago de Compostela, Spain, and they can be seen in ( http://www-gsi.dec.
usc.es/ roberto/TIN2009.html ). Because of the position of the lights inside the
building we will see the performance of our system when the illumination condi-
tions change significantly. This alters the colour aspect of the torsos detected from
the camera located on the robot. Texture is also affected by both illumination and
person's pose changes. The robot used in the tests is a Pioneer 3-DX, equipped
with a Sick LMS 200 laser scanner and a camera, both, the robot and the cam-
era are controlled from a laptop with a 2.4GHz Core 2 Duo processor. The sys-
tem was coded using the OpenCV library [10] and C++ language, our software is
able to process about 10 frames per second (320 240 images). We compared the

performance of the camera discrimination module using both, the classic average
dissimilarity values Eq.4, and our adaptive weighted dissimilarity Eq. 8.

×

Fig. 3. Evolution of the average dissimilarity between the distractors on the distractors
list and the target. We can see the results using the classic average dissimilarity and
the online feature weighting.

In the first video sequence there are two distractors moving close to the tar-
get being followed by the robot. Figure 3 shows how the performance increases
when feature weighting is used. On average there is an increase of around 29% on
target discriminability (the dissimilarity between the target and the distractors
increases from 0.38 to 0.49). Fig.4 illustrates how the weight values are dynam-
ically adapted as the time goes by, this change of the weights is mostly due to
the varying illumination conditions. In general the situation in which there are
several distractors moving close to the target is the riskiest. Luckily, when there
is only one distractor (even thought the distractor might change after several
frames), the weighted dissimilarity increases the separability between it and the
target (regions P 0 , P 3 and P 4 in Fig.3 and Fig.4). It is also noticeable that when