| With the development of multimedia and internet technologies, there is a growing interest in video mining research that is to discover knowledge existing in the video data. The major challenge of video mining is how to bridge the semantic gap between low-level features and high-level semantics, which characterizes the difference between two descriptions of the video data, i.e., the computable visual descriptions and understandable semantic interpretations.;In this dissertation, we proposed a new sports video mining framework where a hybrid generative-discriminative approach is used for multi-level video semantic analysis. A three-layer semantic space is proposed, by which the semantic video analysis is converted into two inter-related statistical inference procedures at different levels. The first is to infer the mid-level semantic keywords from the low-level visual features via generative models, which can serve as building blocks of high-level semantic analysis. The second is to detect high-level semantics from mid-level semantic keywords by using discriminative models, which are of direct interests to users.;Specifically, HMMs-based approaches and CRFs-based approaches are employed in two inference problems respectively. In the first inference problem, to explore multiple co-existent semantic keywords jointly, we developed a multi-channel segmental hidden Markov model (MCSHMM), which integrates both hierarchical and parallel dynamic structures to offer more flexibility and capacity of capturing interactions between multiple Markov chains as well as incorporates the segmental HMM (SHMM) to deal with variable-length observations. In the second inference problems, in order to segment and recognize the high-level semantics, i.e., the game flow, simultaneously, we proposed the auxiliary segmentation conditional random fields (ASCRFs) that can both group a set of plays into different drives and handle possible missing keywords with an additional auxiliary layer.;The use of hybrid generative-discriminative approaches in two different stages is proved to be effective and appropriate for multi-level semantic analysis in sports video. The experimental results from a set of American football video data demonstrate that the proposed framework offers superior results compared with other traditional approaches. |
