Football Players Classification in a Multi-camera Environment - Advanced Concepts for Intelligent Vision Systems - page 150

Fig. 7. Classes Temporal Evolution over a whole match

Table 2. Overall performance of the classifiers with Transformed RGB histograms

MBSAS BCLS K-Means
NO CBTF 72.21% 81.33% 84.65%

CBTF 76.19% 84.24% 87.38%

transformed one (by means of CBTF). Figure 5 shows the original training data
configuration by PCA decomposition. In figure 6 the transformed data are plot-
ted using PCA decomposition technique. As the reader can see the CBTF (trans-
formed data in figure 6) permits to separate the classes without overlapping and
the relative cluster are better spaced.
Overall results of training both in presence/absence of preprocessing based
on CBTF are presented in table 1. As it can be noted, the best overall results
have been reported by using the Transformed RGB features in presence of CBTF
preprocessing. It is probably due to the spectral invariancy introduced by them,
while more sensible features, like simple RGB histograms, perform worse. How-
ever, the perfect separation of clusters has not been obtained for all sequences:
by accurately observing images, in some football matches we noted that some
clusters are really dicult to distinguish even for humans. For example, some-
times one of the goalkeepers was dressed in a very similar way with the referee,
while in another match a goalkeeper was dressed like players of opposite team.
In this case a correct classification based only on spectral information (without
considering the player position in the play field) is really dicult also for human.
In fig. 4 an example of two ambiguous classes is illustrated. Unfortunately, from