Sports Video Semantic Content Analysis

One of the major challenges facing current content-based video analysis and the related applications is the so-called "the Semantic Gap" between the rich high-level semantics that a user desires and the shallowness of the low-level features that the automatic algorithms can extract from the media. In this thesis, we systematically explore the problem of bridging this gap in the sports video.According to domain-specific knowledge of sports video, at first we define those periodical or semi-periodical important semantic parts during the sports programs as "Basic Semantic Unit", abbreviated to "BSU", which include AudioBSU, SceneBSU and EventBSU and so on. Then a general framework based on BSU for sports video semantic content analysis is presented. Within this general framework, we develop the methods of BSUs semantic content analysis that map low-level features to high-level semantics. Finally, the above framework and methods are validated by designing and implementing the Sports video Semantic Content Analysis and Summarization Platform-- SCASP.The main contributions of this thesis are as follows:· We propose a novel unified BSU-based framework for sports video semantic content analysis, which is composed of two parts: the concept model BSUCN (Basic Semantic Unit Composite Network) and the probabilistic technical framework. On one hand, BSUCN defines the relations among BSUs as "BSURelation" and models the semantic content of sports video. To extract semantics from sports video, we convert the video indexing and understanding problem into a pattern classification and recognition problem. On the other hand, the technical framework clarifies the appropriate approach and methodology of this domain. Unlike previous approaches, we want a feasible, general and effective technique for developing those stochastic models rather than fine-tuning signal-based analytical procedures.· We address the method of AudioBSU semantic content analysis based on Gaussian Mixture Model. We model three kinds of AudioBSU in sports video using GMM and approach the AudioBSU semantic content analysis as audio classification and segmentation.· We develop the method of SceneBSU semantic content analysis based on Hidden Mixture Model. We model the statistical temporal relations of views and scenes in sports video using HMM and approach the SceneBSU semantic content analysis as scene classification and segmentation.· We devise the method of EventBSU semantic content analysis based on Bayesian Network. We model the combined relations of low-level evidences in event and approach the EventBSU semantic content analysis as fusion analysis in event detection.· We design and implement SCASP, which gives a sound support to the aboveframework and methods of sports video semantic content analysis.In a word, this thesis provides an in-depth investigation into the concepts, framework and methods of sports video semantic content analysis. The framework and methods are flexible and generic and can therefore be applied to applications such as multimedia management, human-computer interaction and so on.