Interpreting sports tactic based on latent context-free grammar

In this paper, latent context-free grammar (LCFG) is proposed to probabilistically interpret high level tactic concepts in sports video. From domain knowledge, a sports concept typically consists of multiple levels of recursive or non-recursive sub-concepts. Conventional shallow models, e.g. HMMs, have difficulties in characterizing such complex semantics. On the other hand, a comprehensive Bayesian network may require detailed design and parameterization, which is frequently impractical. LCFG is introduced as an extension to stochastic context-free grammar (SCFG), which jointly uses a set of low level discriminative terminals from video analysis and a set of intermediate context-free rules from sports domain knowledge to model the complex athletes´ behaviors and the underlying tactics. The classical `pick-and-roll´ tactic in basketball game is studied in our experimental work. The experimental results demonstrated the rich representation and interpretation powers of LCFG through its probabilistic parsing trees.