Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 151

size of a segment is 20 seconds) is around 6 seconds. Hence, the overall activity
retrieval process took 2 minutes and 20 seconds. The HPaHMM computation time
is a little lower than the HPaSMM one. Computation times are much smaller than
those reported for squash video processing (see Section 3.4). Indeed, kernel approx-
imation of the trajectories is the procedure that took most of the computation time
for squash video processing and it is not needed here.
Moreover, since referees whistles detection allows to decode each segment in-
dependently, it is possible to decode a segment as soon as it ends. When a whistle
is blown and detected, around six seconds (depending on the size of the processed
segment) are necessary to retrieve activity phases in that last segment. Using audio
information hence allows an almost real-time processing of handball trajectories.

5 Conclusions, Extensions and Perspectives

We described here an original hierarchical semi-Markovian framework for sports
understanding purposes that uses trajectories extracted from videos. An upper level
layer is dedicated to modeling of high semantic activity while a lower level layer
is used to process low level feature values. This is a general method that may be
used for many kind of sports, and more specifically for team and racquet sports.
Trajectories of players reconstructed in the court plane are used for activity phases
recognition.
The proposed method has been adapted to two very different sports: squash and
handball. For each of these two sports, a specific representation taking into account
both dynamics and interaction information is proposed and a set of relevant activ-
ity has been defined. The developed HPaSMM method has been tested on two sets
of trajectories corresponding respectively to approximately 10 minutes of a squash
game and of a handball game. Experiments gave very satisfying results, showing
efficiency and easiness of the adaptation of the proposed modeling to very different
sports. Comparison with Markovian modeling highlighted the importance of mod-
eling activity state durations provided by semi-Markovian modeling. Extensions to
more complex model architectures could be handled with larger sets of trajectories.
Thanks to the explicit activity phase modeling, extensions of the proposed
method to other sports (such that soccer, basket-ball, tennis, volley-ball) looks
straightforward. To this aim, at a lower level and depending on the processed sport,
a specific feature value representation would have to be defined. However, it would
still be important to adapt a reduce feature value representation that keeps taking
into account both for individual players movements and for interaction. Moreover,
activity phases definition is depending on the considered sport, and also on the de-
sired degree of automatic understanding.
Based on these modelings, interesting applications would be considered in sev-
eral domains. Firstly, relevant use for the mass media are reachable, including for
example automatic sports video summarization and video on demand requests. Be-
sides, such semantic analysis are of great interest in the sports professional field,
allowing automatic players statistics generation, tactical phases understanding...