Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 147

Fig. 14 Plot of the recognition results obtained by the HPaSMM method. First part of the
trajectory set (6370 images, see Fig. 13) was used for training while the second part (8294
images, see Fig. 13) was kept for recognition test. Ground truth is plotted in red and recog-
nition results are plotted in blue. Numbers on the vertical axis correspond to the numbering
of activity phases described in Section 4.2. The horizontal axis denotes the numbering of the
successive groups of images (with parameter k group value set to 8, see Section 2.2).
segmentation results. With the HPaHMM method and the same training and testing
data sets, a rate of 72.7% has been reached.
Such results can be explained by the lack of training trajectories. For sure, 4 min-
utes for the training set does not seem sufficient to efficiently train such ”complex”
models. Moreover, most of the errors occurs for the offense and defense free-throw
or timeout activity phases. To alleviate these shortcomings, we will now exploit au-
dio information contained in the audio stream. Indeed, using the LOOCV available
when processing audio information will provide us with larger training sets while
helping detection of transitions of free-throw or timeout activity phases.
Hence, in the following, we first present results of referees whistles extraction.
Then, results of activity understanding conducted with the leave-one-out cross vali-
dation method, which enables to consider more training data, are reported.
Results of whistle audio recognition
The method presented in Section 4.3 for recognition of referees whistles gave very
satisfying results. Indeed, 29 referees whistles out of 31 where recognized while no
false detection was detected. Figure 15 illustrates these results, where green lines
correspond to detected whistles and brown ones are undetected whistles.
We now use the referees whistles (29 referees whistles were correctly detected in
the ten-minutes audio stream) in the HPaSMM and HPaHMM methods.