Football Players Classification in a Multi-camera Environment - Advanced Concepts for Intelligent Vision Systems - page 143

system less dependent from particular match conditions (for example the a-priori
knowledge about the team uniforms). Supervised approaches based on spectral
contents are proposed in [12] (based on the analysis of colors in HSI space),
[10], [13]. In [11] the position of each player in the field is integrated to make
the classification more reliable. A recent interesting work working on broadcast
moving images has been proposed in [2]. Moreover in a multi-view context a
Cumulative Brightness Transfer Function (CBTF) is proposed [7] for mapping
color between cameras located at different physical sites, which makes use of
the available color information from a very sparse training set. A bi-directional
mapping approach is used to obtain an accurate similarity measure between
pairs of candidate objects.
All the above works try to solve the problem of player team discrimination
inasupervisedwayandonasinglecamera view, by means of human-machine
interactions for the creation of the reference classes. In this work we investigate
on the usability of unsupervised algorithms for the automatic generation of the
class models from patches coming from different cameras (players and referee).
The proposed work analyzes two main aspects of unsupervised classification: the
selection of the best set of features, and the selection of the best classifier for the
examined application context. Moreover, the problem of different appearance of
players in different views, or in differently lighted regions in the same view, is
analyzed; an approach based on the evaluation of the Cumulative Brightness
Transfer Function (CBTF) [5] with the goal of referring each player appearance
to the same color model is proposed. Several factors, such as varying lighting
conditions during the match, the overall shape similarity among players, time
constraints for real time processing, make a football match a challenging arena
for pattern recognition based on color descriptors. Therefore, this work try to
be a starting point for all researchers that approach the problem of automatic
analysis of football videos.
We started from the players segmentation algorithm proposed in [8]. For
each detected player, different feature set have been tested: in particular, we
have compared performance obtained with RGB histograms, rg normalized his-
tograms and the transformed RGB (standard RGB histogram modified in order
to obtain histogram with zero means and standard deviation equal to one).
Then, three different unsupervised classification algorithms have been imple-
mented and tested. We have chosen a sequential algorithm (MBSAS - Modified
Basic Sequential Algorithm Scheme), a competitive one (BCLS - Basic Compet-
itive Learning Scheme), and a hard-clustering scheme (Isodata, also known as
k-means). All experiments have been performedbothinabsenceandpresence
of the preprocessing based on the CBTF, finalized to mitigate different color
appearance between different sources.
In the rest of the paper, firstly the system overview is summarized (section 2);
then features extraction procedures (section 3) and the Cumulative Brightness
Function are presented (section 4). After, the classification algorithms are briefly
illustrated (section 5). The experimental results obtained on real image sequences