Football Players Classification in a Multi-camera Environment - Advanced Concepts for Intelligent Vision Systems - page 149

Team 2

Team 1

10

Referee

5

0

GK 2

−5

GK 1

−10

20

−15 0

20

15

10

5

0

−5

−10 −20

−15

Fig. 6. PCA decomposition of Training Data after CBTF

Table 1. Reliability of the training procedure

RGB rg T-RGB
No MBSAS 71.24% 86.22% 93.12%
BCLS 77.77% 87.33% 94.78%
CBTF K-Means 81.43% 88.04% 95.31%
Overall 78.99% 87.31% 94.96%
CBTF MBSAS 74.12% 87.56% 94.31%
BCLS 79.38% 89.45% 95.11%
K-Means 85.83% 91.33% 97.17%
Overall 82.13% 89.96% 96.18%

analyzing the results, here we report the processing parameters values for each
algorithm (64-bin histograms have been used).
-

MBSAS :th=0.5

- BCLS : μ =0 . 2, epochs=10000, exit th=0.01

: k=5, exit th=0.01
In the first experiment we have compared the capability of the training proce-
dure to correctly detect the output clusters according to the different feature
sets. For this purpose we carried out 10 experiments on 10 different matches;
for each of them, about 1800 actors (players, goalkeepers and referees) images
have been randomly collected in the training set, and provided to the algorithms.
Note that these images have been acquired by different cameras, so there could
be some differences in light conditions, as well as in color appearance. An exam-
ple image, with four FOVs in which the illumination is variable, could be seen in
fig.3.Beforetostartwithcolor feature extraction as explained in section 3
we have evaluated the different classes configuration (Team one, Team two,
Goalkeeper one, Goalkeeper two and referee) in the original data and in the

- K-means