A Virtual Curtain for the Detection of Humans and Access Control - Advanced Concepts for Intelligent Vision Systems - page 105

G W + H ( i,j,t )or G W × H ( i,j,t )) is the concatenation of a given number L of

consecutive intra-frame signatures.

3.2 Classification
Classification consists in the learning of a function from labeled input data. The
learned function, sometimes referred to as model , is later used to predict the label
of unlabeled data. The purpose of the classifier is to generalize the knowledge
present in the labeled examples to new samples.
When a person crosses the virtual curtain, we reconstruct the time series of
his binary silhouettes and assign a class to it with our classification algorithm.
In this particular application, only two classes can be assigned to a series of
silhouettes:
- “0”, which denotes that a single person has crossed the virtual curtain,
- and “1”, which denotes that more than one person have crossed the virtual
curtain.

Learning and cross-validation. To build a classifier, it is necessary to label
(manually) a large amount of data samples. Part of these labeled samples are
used to train the classifier. They constitute the “learning set”. Remaining labeled
samples are used to evaluate the performances of the classifier; they are part of
the “test set”.
A rule of thumb is to divide the available labeled data in two equal parts: one
to train the model, and the other to test it. With only a few available labeled
data, it may be disadvantageous to ignore half of the labeled data to train
the model. A common solution is then to use cross-validation briefly described
hereafter.
If there are N labeled samples, cross-validation consists in dividing them into
K subsets of equal size. ( K −

1) subsets are used to train the model while the
remaining one is used to test it. This procedure is repeated K times, for each
test set on turn. The final score of the classifier is the average of the K computed
scores. When K = N , this method is called leave-one-out .

Classification tool. There are many classification techniques available. Among
the most popular are nearest neighbors classifiers (KNN), artificial neural net-
works (ANN), (ensemble of) decision trees, and support vector machines (SVM).
In our case, the sets of features extracted from the time series of reconstructed
silhouettes are classified with a support vector machine classifier [18]. An SVM is
a binary classifier that maps its input features into a high-dimensional non-linear
subspace where a linear decision surface is constructed. We used the implemen-
tation provided by libsvm [19] with a radial basis function (RBF) kernel.

4 Results

To evaluate the performances of our algorithm, the B.E.A. company has pro-
vided us 349 labeled sequences of a single person (class “0”) and 517 sequences