Information Technology Reference
In-Depth Information
sequences of the same team in a same game. It would be an add to test the proposed
method on trajectory data from other game and other teams.
The first experiments processed use of the HPaSMM and HPaHMM methods
without considering audio information. A part of the handball players trajectories
is used for training the HPaSMM and HPaHMM models. It includes 6370 images,
i.e. , more than 4 minutes of video. A second part of the trajectory set is then used
for testing and comprises 8294 images, i.e. , little less than 6 minutes of video.
Figure 13 presents test and training data sets.
Fig. 13 Left: training set of handball players trajectory (6370 frames). Right: testing set of
handball players trajectory (8294 frames).
We have also tested the proposed HPaSMM method while taking into account
audio detection of the referees whistles. The LOOCV validation method was then
considered, allowing larger training data sets while keeping non-overlapping test
and training data sets (see Section 4.3).
Observations from other video streams have also been included to train the up-
per level state transition matrix A and the GMMs modeling the state duration sd i .
Indeed, trajectories are not required to train this subset of parameters. Hence, other
handball videos were used and manually segmented into the eight considered activ-
ity, hence providing additional information in term of transitions between activity
phases and of durations of activity phases. These handball activity data have been
extracted from 2008 Beijing Olympic Games handball final. The videos are avail-
able online [5].
In the sequel, we will evaluate the performance of the method as the ratio of cor-
rectly classified images (with respect to the eight defined activity phases) and the
total number of processed images. To this end, ground truth on the entire set of tra-
jectories is exploited. All the reported results were obtained using k group parameter
value equal to 8 (as defined in Section 2.2).
4.5
Results
We report here results corresponding to the experiments described in the previous
section. First, when using training and testing data sets of respectively 6370 and
8294 frames with the HPaSMM method, a rate of 76.1% correct recognition was
obtained. Results are plotted in Figure 14 which contains the ground truth and the
 
Search WWH ::




Custom Search