Online Feature Weighting for Human Discrimination in a Person Following Robot - Foundations on Natural and Artificial Computation - page 225

distances we can compute the average dissimilarity amongst each torso and the
target model:

n

1
n

dissimilarity = d f ∈ [0 , 1] (4)

f =1

n is the number of features used to describe our target.
Every torso is compared with the target model to see if there is one that is
similar enough to be considered the target (we use a threshold value thr 1 =0 . 4
to do this). Nevertheless, if this dissimilarity is below a second and more restric-
tive threshold ( thr 2 =0 . 2), the target model will be updated using the torso
identified as the target. This dual-threshold strategy avoids the pollution of the
target histograms with wrong detections of the target. However, the dissimilar-
ity measure described before (Eq. 4) is not yet robust enough to cope with real
world conditions, such as strong illumination reflections, shadows, occlusions,
and situations where the algorithm has to discriminate among similar torsos.
Because of this, we have designed an adaptive weighting process to enhance the
discriminability between the target and the distractors, this process is described
in the following section.

2.3 Online Feature Weighting

Online feature weighting is the process of assigning high weights to the features
that show a high discriminability between the target and the distractors. This is
very useful when target and distractors share some common feature distributions.
We can think, for example, in a target and several distractors wearing similar
colour clothes, in this case it would be more useful to focus the dissimilarity
measure on the texture features rather than in colour. This is what we try
to achieve through the algorithm described in this section: an on-line adaptive
process that changes the weights of the features trying to maximize the impact
of the most discriminative ones.
As shown in Fig. 2, the feature weighting process consists on two tasks: a)
building and updating a distractor list, and b) updating the values of the feature
weights. The first task consists on building a distractor list containing all those
torsos that have not been considered as the target. The distractors list is built
and updated at every frame according to the following rules:

1. If a torso can be classified as the target, the rest of the torsos that have been
detected in the same frame will be placed on the distractor list provided that
they do not overlap in the image with the torso corresponding to the target.
2. If there is not a torso that can be classified as the target, those with a
dissimilarity value higher than 0.5 will be put on the distractor list.
3. The list has a size limit of ten torsos. When the limit is reached, the oldest
torsos will be removed. This size limit of the list is set to consider only the
most recently seen torsos. A larger distractor list would save torsos which will
not be seen again in a short period of time, thus reducing the performance
of the feature weighting process.