Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 143

- “offense free-throw or timeout”: activity phase 3,

- “counterattack, fast break”: activity phase 4,

- “returning, preventing from fast break”: activity phase 5,

- “slowly returning in defense”: activity phase 6,

- “defense”: activity phase 7,

- “defense free-throw or timeout”: activity phase 8.

Such a variety of activity provides a quite complete picture of the possible hand-
ball activities, providing an exhaustive representation for efficient understanding of
handball videos. In the illustration in Figure 10, the upper level layer is surrounded
in red and is composed of eight upper level states.

Lower layer: feature modelings using parallel hidden Markov models

The five feature values that describes both dynamics of the trajectories and inter-
actions between trajectories, i.e. , d GF , t , d F , t , d intramin , t , d intramean , t ,and d intramax , t are
used to characterize the eight phases of activity S i . For each upper level state, the
five corresponding feature vectors are modeled using a 5-layer PaHMMs. Figure 10
presents the lower level layer surrounded in green.

4.3 Integrating Audio Information: Recognition of Referees

Whistles

To facilitate recognition of activity phases and more precisely their transitions, audio
information is taken into account. Indeed, audio data contains an important infor-
mation to specify transitions between activities: referees whistles.
To retrieve referees whistles instances, the audio stream is processed using two
free access softwares: Spro and Audioseg, available online [26, 2]. Spro provides
a description of whistles request and of the processed audio stream contained in
the handball video. This characterization is based upon cepstral coefficients of mel
frequency [15] computed in successive time intervals defined by a sliding win-
dow. Audioseg then performs recognition of the request within the audio stream by
comparing the coefficients of each audio intervals using a Dynamic Time Warping
procedure [16].
Detection of referees whistles is integrated in the HPaSMM method in a simple
way: each whistle corresponds to an activity phase change in the model. Hence,
in the Viterbi decoding algorithm, the hypothesis is made that the current activity
phase is stopped each times a whistle is detected and then another phase begins.
Hence, a partition of the observed actions into successive segments Seg k is given
by referees whistles, where a segment is included between two referees whistles.
Each segment can then be decoded separately using the Viterbi algorithm. Since
each whistle corresponds to a change of activity, decoding of successive video seg-
ments is simply made by specifying that first activity phase of a segment Seg l + 1 has
to be different from the last activity phase found for the previous segment Seg l .