Abstract
Automated video annotation is a topic of considerable interest in computer vision due to its applications in video search, object based video encoding and enhanced broadcast content. The domain of sport broadcasting is, in particular, the subject of current research attention due to its fixed, rule governed, content. This research work aims to develop, analyze and demonstrate novel methodologies that can be useful in the context of adaptive and automated video annotation systems. In this thesis, we present methodologies for addressing the problems of anomaly detection, rule adaptation and rule induction for court based sports such as tennis and badminton. We first introduce an HMM induction strategy for a court-model based method that uses the court structure in the form of a lattice for two related modalities of singles and doubles tennis to tackle the problems of anomaly detection and rectification. We also introduce another anomaly detection methodology that is based on the disparity between the low-level vision based classifiers and the high-level contextual classifier. Another approach to address the problem of rule adaptation is also proposed that employs Convex hulling of the anomalous states. We also investigate a number of novel hierarchical HMM generating methods for stochastic induction of game rules. These methodologies include, Cartesian product Label-based Hierarchical Bottom-up Clustering (CLHBC) that employs prior information within the label structures. A new constrained variant of the classical Chinese Restaurant Process (CRP) is also introduced that is relevant to sports games. We also propose two hybrid methodologies in this context and a comparative analysis is made against the flat Markov model. We also show that these methods are also generalizable to other rule based environments.