Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 138

Hence, it is difficult to visually determine if the two squash players are in a “rally”
phase or in a “passive” phase in the video. Players movements may be very reduced
both in the “rally” phase and in the “passive” phase. Indeed, even during the video
segments of “rally” activity, there are periods during which players are almost stat-
ics, so that it looks more like a “passive” phase. To visually proceed a temporal
segmentation, a simpler way seems to focus on the relative distance evolution be-
tween players ( i.e. , the evolution of the feature d ). Hence, such experiments are of
interest and may justify the chosen activity characterization that relies on players
motions and on their interaction.
In the sequel, we will evaluate the performance of the method as the ratio of
correctly classified images (with respect to the “rally” and “passive” activity phases)
and the total number of processed images. To this end, ground truth on the entire set
of trajectories is exploited. All the reported results were obtained using k group and h
parameter values respectively equals to 8 and 1 (as defined in Section 2.2).

3.4 Results

Hence, the first 7422 frames of the video (and more precisely the corresponding
trajectories) where used to learn squash HPaSMM showed in Fig. 5. This step re-
sults in the computation of the parameter set A , φ , ψ }

defined in Section 2.3.
Fig. 8 shows the training result when fitting a GMM on the duration state distribu-
tion of respectively the “rally” and “passive” upper level state.
A 89.2% of correct classified images has been reached on the 8086 testing images
with the squash HPaSMM. Corresponding results are presented in the upper part
of Figure 9. Using the comparison Markovian method ( i.e. , the squash HPaHMM
model developed in Section 2.5), a 88% correct recognition rate was obtained. Cor-
responding results are shown in lower part of figure 9.

θ = {

Fig. 8 Left: duration state density modeling using GMM for “rally” upper level state. The
x-axis corresponds to the observed state durations (crosses indicate observed state durations).
Left : duration state density modeling using GMM for “passive” upper level state.