Football Players Classification in a Multi-camera Environment - Advanced Concepts for Intelligent Vision Systems - page 148

(a) Goal- (b) Referee

keeper

Fig. 4. Example of hard-distinguish players

GK 2

0.1

Team2

0.05

GK 1 0

−0.05

Team 1

−0.1

−0.15

Referee

−0.2

−0.2

−0.25

0

0.2

−0.3

0.4

−0.2

−0.15

−0.1

−0.05 0.6

0

0.05

0.1

0.15

Fig. 5. PCA decomposition of Training Data

light conditions. The training set has been created by collecting a great number
of player feature vectors, randomly selected (many times repeated) from real
football images, with the care of including players positioned in different parts
of the play field (to ensure the inclusion of goalkeepers and lineman referees).
This feature set has been used during the training phase of the classifiers; each
cluster has been represented by means of a feature vector ('representative' of the
cluster). Then, at runtime each segmented player is provided to the classifier for
the test phase. However, in this kind of applications, each game is a different
case, and overall results could be misleading. For example, in a match we can
have well-contrasted uniforms, with well separated classes, while in another one
the classes could overlap in the feature space. For this reason in the following we
present results obtained both on several matches (for testing the training phase)
and on a single, random selected, match (for the test phase evaluation). Before