Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 140

However, HPaSMM provides an imperfect segmentation. Indeed, 14 points instead
13 were detected, hence retrieving one additional (non-existing) point. It highlights
the contribution given by the semi-Markov upper level modeling, validating the hy-
pothesis that state duration in upper level state does not follow a geometrical law
and needs a more adapted modeling. This is illustrates in figure 9, were the purple
surrounded region contains the false “rally” detection.
Results below 70% of good phase segmentation obtained when not considering
the d feature value show the decisive inherent information of the temporal trends of
the distance ( i.e. , of the interactions). So, single players motions are not sufficient to
understand the observed activities. Results obtained using only the interaction fea-
ture vectors D S i shows that the interactions between players is a crucial information
for activity phase retrieving. Moreover, comparison with HPaSMM method when
using D S i as well as V S i

1 and V S i 2 feature vectors shows that exploiting dynamics
and shapes of the single trajectories in addition to interactions helps getting a better
recognition. These results validates the hypothesis made by our modeling that both
single players dynamics and their interactions are to be considered to understand
activities in sports such that squash.

Computation time

We give here computation time results using an Intel Pentium Centrino 1.86 GHz
processor. Learning the set of HPaSMM parameter A , φ , ψ } with the 7422

frames training set took around 1 minute long. Computation time for the recognition
stage on the 8086 frames was around three minutes. The HPaHMM computation
time is slightly lower than the HPaSMM one.
Nevertheless, when considering HPaSMM only with the feature vectors D S i used
for activity representation (results skipping from 89.2% to 86.8%), computation
times become much lower. Indeed, most of the squash HPaSMM computation time
is spent computing the continuous representation necessary to get the ˙

θ = {

values.
Hence, when considering only the D S i feature vectors, no kernel approximation is
further needed and the computation times are much smaller (divided by around 10
times).

γ

4 HPaSMM Handball Activity Recognition

In order to show the efficiency of the HPaSMM framework for sports video under-
standing issues, an application of HPaSMM to handball video is here presented. It is
a more complex use of the HPaSMM model than with squash videos, since handball
teams are formed with seven players and the diversity of activities is higher.

4.1 Handball Invariant Feature Representation

To process semantic segmentation of handball videos, motions in the court plane
( i.e. , the trajectories) of the players of a single team are considered. Similarly to