Football Players Classification in a Multi-camera Environment - Advanced Concepts for Intelligent Vision Systems - page 144

acquired during football matches of the Italian Serie A are described in section
6. Finally, conclusions and future works are reported in section 7.

2Sy emO rvew

The multi-camera environment consists of a real system installed in the ”Friuli”
stadium situated in Udine (Italy). This prototype permits to detect automat-
ically ”offside” during the football match [15]. The system is composed by six
high resolution (Full HD) cameras (labeled as FG i ,where i indicates the i-th
cameras) placed on the two sides of the pitch. This location assures double cov-
erage of almost all the areas by either adjacent or opposite cameras. In figure
1 the location of the cameras is shown. The acquired images are transferred to
six processing nodes by fiber optic cables. The acquisition process is guided by
a central trigger generator that guarantees synchronized acquisition between all
the cameras. Each node, using two hyper-threading processor, records all the
images of the match on its internal storage unit, displays the acquired images
and, simultaneously, processes them with parallel threads, in an asynchronous
way with respect to the other nodes. The six processing nodes, are connected to a
central node, which has having the supervisor function. It synchronizes the data
coming from nodes and performing high level processing. The figure 2 shows the
six images acquired from the six nodes linked to the cameras located around the
pitch (see figure 1). Each nodes uses a motion segmentation algorithm [8] based
on statistical background subtraction. Information relative to moving objects
are the used to perform human blob detection. The player blobs represent the
starting point of the classification step. We have evaluated the performance of
different combination of unsupervised classifier and color feature applied in a
multi-camera environment.

3 Feature Selection

In order to separate players in different classes, they should be represented by
a features vector able to emphasize both intra-class analogies, and inter-class
differences. Moreover, the selected features should be as well scale invariant (im-
ages of players could have different size according to the geometry of acquisition
sensors and their position in the field), rotation invariant (usually players are
standing, but sometimes they can appear slanted on the field), and also quickly
extractable (real time processing is often a fundamental requisite for sport analy-
sis applications). Starting from these requirements, we have tested three different
feature sets that satisfy the above mentioned conditions:

RGB histograms: the RGB histogram is a combination of three 1-dimensional

histograms based on the R,G and B channels of RGB color space.
rg histograms: in the normalized histograms the chromaticity components r

and g describe the color information in the image; it is robust to light vari-
ations in luminannce;