Multi-channel segmental hidden markov models for sports video mining
Yi Ding, Guoliang Fan · 2008
We present a new multi-channel segmental hidden Markov model (MCSHMM) for sports video mining that is a unique probabilistic graphical model with two advantages. One is the integration of both hierarchical and parallel structures that offer more flexibility and capacity of capturing the interaction between multiple Markov chains. The other is the incorporation of the segmental HMM that represents a variable-length sequence of observations. Especially, we develop a maximum a posteriori (MAP) estimator to optimize model structures and model parameters simultaneously. The proposed MCSHMM is used for American football video analysis, where two semantics structures, play types and camera views, are involved. The experiment shows that the MCSHMM outperforms previous HMM-based approaches.