Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 145

Fig. 12 Complete set of available trajectories of the seven players belonging to a same hand-
ball team (each trajectory is plotted in a given color and corresponds to a different player). It
corresponds to about ten-minutes of video (more precisely, 14664 frames).
players belonging to a same handball team. They correspond to approximately ten
minutes of video (14664 frames).
The database also provides a ground truth annotation of the video within several
activities. The set of activities considered here is slightly different from the eight
activities we have chosen. Indeed, thanks to ground truth, activities denoted by “7-
meter throw” and “Jump ball at the court center” are observed only once in the pro-
cessed ten minutes video. Hence, in order to generate ground truth, we were forced
to segment these events using the eight defined activities. With a larger amount of
trajectories, such events could have been properly modeled. This stresses the im-
portance of having a sufficient amount of training data adapted with the considered
modeling.
Likewise, the event “Back in the court center” (after conceding a goal) is ob-
served only twice in the whole video. This activity would also have required the
setting-up of an additional upper state. Again, it was not possible to efficiently model
it due to the available trajectory data sets. In order to produce corresponding ground
truth, we had to resort to the activities “defense free-throw or timeout” and “slowly
going into attack” which depict the best a similar motion content.
Hence, we used the available annotation to construct our activity ground-truth
with the eight activity phases defined in Section 4.2. Provided ground truth also
allowed us to validate the hypothesis that every whistle induces a change in the
activity phases.
The experiments carried out to test the proposed HPaSMM and HPaHMM meth-
ods for trajectory-based handball activity recognition are now described.
In our experiments, data sets respectively used for training and testing the pro-
posed algorithm correspond to distinct periods of the ten minutes handball video.
Hence, no overlapping data were used for training and testing. However, due to
the available trajectory data sets, training and testing sets of trajectories are video