Online Feature Weighting for Human Discrimination in a Person Following Robot - Foundations on Natural and Artificial Computation - page 228

Fig. 4. Evolution of feature weights during first video sequence

class separability is very low, one or two features are chosen while the rest of
them are discarded (region P 1 and P 2 in Fig.3 and Fig.4). Table 1 shows the
confusion matrix which shows the recognition ratio, and the number of false
negatives and positives. The results are shown when using our proposal and the
classic average weighting. In order to get the confusion matrix we processed all
the video frames apart from those where the target was standing in front of
the robot without moving, or when there were no illumination changes or no
distractors. Despite of the fact that the target recognition seems to be worse
with our system than with the classic weighting approach, the number of false
positives is clearly lower (situations in which the distractor is misclassified as
target). This is something relevant since missing the target during some few
frames will not alter the behaviour of the robot too much thanks to the laser
module, nevertheless mistaking the target for the distractor would cause an
important misbehaviour on the robot.

Table 1. Confusion matrix of the first 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 215(95,5%) 10(4.5%) Target 189(84%) 36(16%)

Distractor 16(2,8%) 544(97%) Distractor 0(0%) 560(100%)

Regarding the second video sequence, it shows the movement of the target
and one distractor. Fig. 5 shows the increase of the performance when feature
weighting is used. In this second case the dissimilarity is increased, on average,
from 0.35 to 0.58 (66% increase on target discriminability). Neither the target nor
the distractor wear vivid colour clothes, therefore texture features and lightness
are expected to be the main discriminators. Figure 6 shows that the results
confirm what was expected: two of the total number of features represent 80%
of the total sum of weights during most of the video sequence. Analysing Figure 6
we can notice that there is a small time interval at the beginning (from T 0 to T 1 ),
in which the weight values are similar to each other and the dissimilarity between
target and distractor is not very high, this is due to the illumination conditions