Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 144

Setting-up of a leave-one-out cross validation method

The first advantage in taking into account referees whistles is to help recognition of
some transitions between activity phases. Moreover, it facilitates establishment of a
leave-one-out cross validation (LOOCV) test method. Indeed, the set of trajectory
is “cut” into N w +

1 segments (defined by N w detected whistles). Viterbi decoding
of one segment is computed independently from the other segments, the only infor-
mation to integrate being that the first activity phase of a segment Seg k + 1 has to be
different of the last activity phase of the segment Seg k . Hence, a LOOCV method
is exploited. Each segment is decoded independently while the N w other segments
are used for training HPaSMM as well as HPaHMM models. In the following, this
procedure will allow the use of larger training sets.

4.4 Considered Data Sets and Experiments

The proposed method has been tested on a set of trajectories of handball players
belonging to a same team. The trajectory database is available online [3]. To extract
these trajectories, two bird-eye view cameras were used, one above each half of the
court plane. To this aim, a modified color-histogram-based condensation tracking
method was computed on the video sequences [14, 21]. Database providers super-
vised the tracking procedure by correcting errors that appeared during the process.
An appropriate calibration step was also introduced to map the image coordinates in
the court plane and to compensate for the observed radial distortion. An estimation
of the error on the players coordinates in the court between 0.3 and 0.5 m has been
obtained [20]. For each player, 25 coordinate data (one per frame) per second are
available. To reduce the tracking jitter while preserving the measurement accuracy,
a Gaussian smoothing was finally applied.
Three images corresponding to a same given time instant of a handball match
are provided in Figure 11. The two images on the left were extracted from videos
acquired by bird-eye view cameras. The right one is a frame of a video shot by a
hand-held camera. Figure 12 contains the entire set of trajectories of the handball

Fig. 11 Three images of a same instant in three videos of a given handball action. The two
left images have been extracted from two bird-eye view videos, the third one from a hand-held
camera.