Online Feature Weighting for Human Discrimination in a Person Following Robot - Foundations on Natural and Artificial Computation - page 229

and pose of the target. Nevertheless this situation changes soon, achieving a
small set of prominent features that increase the separability amongst target
and distractor. As we can see in Table 2, the recognition ratio is clearly superior
with our system than with the classic average weighting, and once again the
false positive ratio has been reduced.

Fig. 5. Evolution of the average dissimilarity between distractors included in the dis-
tractors list and the target during the second video sequence.

Table 2. Confusion matrix of the second video sequence. Results are shown using the
classic approach (Left), and using our proposed weighting of the feature space (Right).

Actual \ Classif. Target Distractor Actual \ Classif. Target Distractor

Target 139(80.4%) 34(19.6%) Target 142(99.3%) 1(0.6%)

Distractor 8(3,6%) 212(96,3%) Distractor 3(1.3%) 217(98.6%)

These results confirm that our system is capable of adapting to dicult con-
ditions through an on-line weighting process able to maximize the dissimilarity
target-distractor. We have also tested the person-following controller operat-
ing on the real robot during several 10 minute walks around the corridors and
the hall of our department. The robot had to follow the target when both the
corridors and the premises of the building were usually crowded with students
walking around. The robot's maximum speed was set to 1 m/s, thus allowing
the target to walk at normal speed. The robot was able to avoid collisions with
the environment thanks to a potential field method implemented on the robot
controller. The robot was able to follow the target keeping a distance that ranged
from 0.5 meters to 6 meters, although the average distance between the robot
and the target was two meters.

4 Conclusions

We have described a person-following behaviour which combines a laser based
tracker, with the discrimination power of a camera. The camera uses a novel
online feature weighting method. This method evaluates the discriminability of