Probabilistic graphical models for pattern recognition and optical motion capture tracking

This dissertation discusses several probabilistic graphical models which address problems in pattern recognition and optical motion capture tracking. It first introduces the augmented hidden Markov model with equivalence classes (AHMM+EC), which provides a unifying framework for a large number of existing state-based probabilistic generative models. For example, the AHMM+EC can be used to represent regular hidden Markov models (HMMs), terminating HMMs, hierarchical HMMs, semi-Markov models, reduced-parameter models, Markov models of order larger than 1, and conceptual combinations thereof. The advantage of such a unifying framework is that an algorithm specified or implemented for the AHMM+EC is immediately applicable to any of the models it can represent. The dissertation then shows how the AHMM can be extended into the semantic network model (SNM), in which states of probabilistic models can be marked as semantic states. States are marked when they carry some special meaning to the application, for example the beginning or end of a gesture. Defining semantic states allows formulating and solving problems specifically related to semantic states, which is shown useful in segmentation of an unknown observation sequence, event-driven application frameworks, on-line learning, and finding multiple likely explanations of the data. Both the AHMM and SNM models and related algorithms have been implemented in the open source AME Patterns library, which is presented in a broad overview, and with an accompanying analysis of various tradeoffs and design decisions involved. Finally, the dissertation presents a comprehensive method aimed at making optical motion capture more robust and less time consuming. The first part is an autonomous algorithm for the real-time creation of a moving subject's kinematic model from optical motion capture data and with no a priori information. The second part shows how the automatically built kinematic model can be matched to known persistent models to provide consistent labeling of its elements. The net effect is that the tedious subject calibration phase typically associated with motion capture is completely eliminated.