Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 148

Fig. 15 Results of referees whistles recognition obtained on the ten minutes of audio stream.
Audio stream is plotted in blue, detection threshold is in red. The 29 detected whistles are
represented by green lines while undetected ones are plotted with brown lines.
Results using audio information and LOOCV method for HPaSMM
When the LOOCV method is applied to the 30 segments of handball video trajec-
tories, a performance of 89.8% correct activity recognition was reached. However,
recognition errors mainly occur when processing events “7-meter throw” and “Jump
ball at the court center” (see Section 4.4). Then, we omitted these two activities
( i.e. , 5 of the 30 video segments). 87% of the whole ten-minutes video was then
concerned, that is more than 8 minutes 30 seconds, and a correct recognition rate of
92.2% was obtained.
The complete set of activity phase has been correctly recovered, while the 7.8%
of errors corresponds to time-lags at the transitions between activities. Most of these
time-lags are around one or two seconds long. Hence, as soon as the training data
set is adapted to the activity architecture modeled in the upper layer of the model,
the method supplies a quite satisfying understanding of the observed video.
Figure 16 and Table 1 illustrate the recognition results obtained with the HPaSMM
after discarding the five segments corresponding to events “7-meter throw” and
“Jump ball at the court center”. Table 1 contains, for each activity phase, the correct
recognition rates.