Football Players Classification in a Multi-camera Environment - Advanced Concepts for Intelligent Vision Systems - page 151

Table 3. Temporal analysis of performance of the classifiers

MBSAS BCLS K-Means

0' - 15' 76.77% 85.34% 89.21%

16' - 30' 77.73% 84.14% 87.19%

31' - 45' 76.11% 84.72% 86.96%

46' - 60' 73.96% 81.62% 86.41%

61' - 75' 74.85% 80.81% 84.55%

76' - 90' 72.17% 79.92% 83.18%

the experience collected in our experiments in the last years, after viewing sev-
eral games, we can assert that this situation (referees and goalkeepers dressed in
similar way) is almost common in football games, and it drives our efforts into
the direction of introducing a check of the player relative positions to make the
classification more robust.
Starting from the results of these experiments, that demonstrates the better
performance carried out by using the Transformed RGB histograms, we concen-
trate our efforts in order to detect the best unsupervised classifier (using trans-
formed RGB as features set). In the second experiment we compared the three
unsupervised classifiers during the test phase, i.e. we evaluated their capability
to properly classify each actor according to the previously detected classes. In
table 2 the overall performance obtained in the test phase are presented. Again,
experiments have been performed both in absence/presence of CBTF prepro-
cessing. We can note that K-means based approach seems to outperform the
other ones, with a classification rate over then 87%.
In table 3 the results of overall classification as a function of the time are
shown. These results coming from a new experiment: for a single match a ground
truth was created by considering patches of players at a fixed time instants. In
particular we considered 1800 patches (from six cameras after CBTF transform)
extracted every 15 minutes, for a total of 1800*6=10800 patches. As evident all
the classification performances are more reliable in the first minutes and then
they decrease (not strongly) along the time. The temporal variation of clus-
ter configuration has to be expected during the football match, in particular
in outdoor contexts. The great duration of the event (90 minutes plus interval)
is accomplished by variation in light conditions. An observation about our ex-
periments needs to be remarked: we trained the classifier before the kick off,
during the pre-match operations. This training remain unchanged for all the
match. Probably the effects of class configuration changes could be mitigated if
the training was carried out again at the beginning of second half. However, it
is not the best practical solution, it could be unpracticable in real time applica-
tions; moreover the variation could be sudden (switch on/off of artificial lights),
in an arbitrary instant, so it cannot be forecasted. In figure 7 is plotted the
class configuration at the begin of the match, immediately after the half time
interval, and at the end of the match. As evident, some classes greatly changed,
while others changed in a less evident way. However at the end of the match the
clusters are closer and this confirms the results obtained in table 3.