Building an effective high-level interpretation system using (hierarchical) hidden Markov models

The goal of Human Action Recognition (HAR) is to generate interpretation in a high conceptual level from the received observation data. We take the domain of soccer game as our focus in this thesis. The HAR problem is complicated by three main characteristics. First, the decision-making process of soccer players are inaccessible and nondeterministic. Second, besides explaining individual behaviors, multiple perspectives at different levels of granularity are also needed for a better understanding. Last, the interpretation system of soccer game should also take into account the interactions between different players.;Our goal is to build a system for soccer games that can extract the hierarchy of observed behaviors and inter-person interactions. Several tasks will benefit from this research, including developing autonomous commentator systems, producing strategy coaching analysis, and providing solid foundations for decision-making.;We use Hidden Markov Model (HMM) and Hierarchical Hidden Markov Model (HHMM) as our framework. HMM is a powerful probabilistic model dealing with uncertainties for temporal data by introducing hidden states, which is a good tool to model the players' inner transitions during the decision-making process. HHMM is a generalization of HMM, which allows us to obtain multiple levels of abstraction.;The contributions of this thesis are as follows. A thorough and comprehensive literature review about HMM and HHMM is presented first. The problem of building an effective high-level interpretation system for soccer is defined. A hierarchical framework from modeling players' primitive actions to inter-person interactions is proposed. Patterns of atomic actions and two-person interactions are investigated, of which HMM is used to model the primitive actions and HHMM is used to model the hierarchy of action sequences and multiple-player interactions. In the end, experimental results are analyzed and discussed. These results will lead to a better understanding of HAR theoretically and empirically.