Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 139

Fig. 9 Up: Ground truth is plotted in red. Result of the HPaSMM processed temporal seg-
mentation is superimposed in blue. The x-axis corresponds to the frame index. The “1” and
“2” values respectively correspond to the “passive” and “rally” phases. Down: Ground truth
is plotted in red. Result of the HPaHMM processed temporal segmentation is superimposed
in green. The x-axis corresponds to the frame index. The “1” and “2” values respectively cor-
respond to the “passive” and “rally” phases. The purple surrounded region contains the main
difference between results using HPaSMM and HPaHMM.

Experimentations have also been performed without taking into account the fea-
ture vectors D S i describing the evolution of the distance between the two players.
Hence, only V S i 1 and V S i 2 feature vectors were here considered. Recognition results,
for the same experimental training and testing sets, fall from 89.2% to less than
70%.
Finally, the HPaSMM method has been tested based only on interactions. This
means that activity phases characterization only relied on the distance feature vec-
tors D S i , and not on feature vectors V S 1 et V S 2 . Recognition results using only inter-
actions between players reaches a 86.8% positive rate. This is a little fewer than the
89.2% rate obtained when considering interactions as well as players movements.
A recognition rate of 89.2% seems a satisfying results. Moreover, the HPaSMM
algorithm retrieved the exact number of activity phases. The rate of 10.8% recog-
nition errors only corresponds to lags between the detected phase changes. Hence,
every transition between “rally” and “passive” activity phases have been detected,
retrieving the exact number of played points ( i.e. , 13 points).
Comparison between HPaSMM and HPaHMM (with state durations not modeled
by GMMs, i.e. state duration in state i following a geometric law in regards of a ii )
is also of interest. Most of the time, HPaHMM gave a quite similar segmentation.