Understanding Sports Video Using Players Trajectories - Intelligent Video Event Analysis and Understanding - page 126

here is to be able to detect activities as well as transitions between activities. Such
an understanding is now of great interest in the computer vision field. It is motivated
by several applications in various domains: video surveillance, sports video index-
ing and exploitation, video on demand. . . Trajectories of mobile video objects are
now available with tracking systems and are prone to be exploited for video under-
standing. Indeed, trajectories provide a high level description of dynamical contents
observed in videos.
Several approaches have been proposed to exploit mobile object trajectories for
content-based video analysis [23, 1, 12]. Gunsel et al. developed a video indexing
framework based on the analysis of “video objects” [7]. Their method relies both
on the dynamics of the tracked objects and on the interactions between the corre-
sponding trajectories. Based on such interactions, other algorithms were developed
to handle video content understanding. Let us mention a system that models inter-
actions between moving entities in a video surveillance context, relying on Coupled
Hidden Markov Models, that was proposed by Oliver et al. [19]. Other contribu-
tions describe trajectory-based schemes for activity recognition using the definition
of scenarios within multi-agent Semi-Markov Chains (SMCs) [11, 17, 18].
In practice, however, due to the extreme variety of observable video content,
no practical extensive application of these methods has yet been reported. That's
why, in the sequel, we will concentrate on the analysis of video trajectories in the
particular context of sports videos. Indeed, such activities are supervised by a set of
rules while occurring in apriori known (most likely closed) spaces, providing hence
a suitable experimentation field for semantic video understanding methods [13].
Moreover, since most sports (tennis, soccer, rugby, handball...) take place on a 2D
ground plane, players movements are heavily characterized by their 2D trajectories
in the court plane. Such trajectories, that are now extractable using computer vision
tracking tools from the existing and abundant literature, hence provide rich semantic
information on the observed sports videos.
More precisely thus, several approaches relying on video trajectories have been
provided for sports video activity recognition [9]. However, such a method only
provide classification of activities in video shots into two categories: “normal” and
“unexpected” events. One trajectory-based framework, still, produces a semantic
segmentation of sports video into activities. It was proposed by Per s eetal.andis
dedicated to basketball video analysis method [22]. Here, a first process segments
tracked players trajectories into three different classes of basketball activities (of-
fense, defense and time-out), relying on Gaussian mixtures and an EM algorithm
trained on manually labeled sequences. Then, based on a partition of the court, a
second stage achieves a template-based activity recognition of the offense video
segments into three different classes of basketball play: screen , move and player
formation . If not straightforward, extending such a method to other sports seems
conceivable but would most certainly require some further investigations. On the
contrary, in the following of this chapter, our aim is to provide an already general
trajectory-based framework which can easily be extended and applied to most team
and racquet sports.