Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 141

squash feature values (see Section 3), the choice of the feature representation of
handball team activities is based upon the single dynamics of players as well as
their interactions. Three main ideas led the choice of the feature values:
- to consider the global dynamics of fielders ( i.e. , any handball player except the

goalkeeper),

- to exploit the specific role and location of the goalkeeper,

- to keep a limited number of feature values.

Invariant feature values for players motion characterization

To characterize players movements and to handle the first and the third ideas defined
above, i.e. to take into account the global dynamics of fielders within a limited num-
ber of feature values, three feature values are computed at each time instant t .These
feature values are the min, mean and max values of distance between successive po-
sitions ( i.e. , the distance covered between t

1et t ) of the six fielders, respectively
denoted by d intramin , t , d intramean , t and d intramax , t . These three feature values gives a
reduce yet important and global information on the dynamics of fielders. Computed
distances correspond to Euclidean distances between trajectories in the court plane.
The values d intramin , t for an activity S i are gathered in the vector D S i

−

intramin :

D S i
intramin

=[ d intramin , 1 , ..., d intramin , n i − 1 , d intramin , n i ] ,

where n i is the size of the processed trajectories in activity S i . The same holds for
the two other feature values d intramean , t and d intramax , t to produce vectors D intramean
and D intramax .

Invariant feature values for players interaction characterization

To characterize players interaction while handling the second and the third ideas
outlined in the introduction of this section, i.e. to exploit the specific role of the
goalkeeper within a limited number of feature values, two feature values are com-
puted at each time instant t :

- the mean distance between the goalkeeper and the six fielders d GC , t ,
- the mean distance between the six fielders d C , t .
The values d GC , t for an activity S i are gathered in the vector D S GC :

D S GC =[

d GC , 1 , ..., d GC , n − 1 , d GC , n ] ,

where n i is the size of the processed trajectories in activity S i . The same holds for
the feature values d C , t to produce feature vectors D S C .
It is important to precise that, on the contrary to squash trajectories feature rep-
resentation, no kernel approximation is here needed. Indeed, the five considered
feature values do not require first or second differential values and are directly com-
puted on the row trajectory data, i.e. ,on T k = { ( x 1 , k , y 1 , k ) , .., ( x l k , k , y l k , k ) } .