This thesis focuses upon the use of Multiple-Sequence training procedures for Hidden Markov Models in gesture recognition applications.
HMM Training is of great interest since there are many important applications in which HMMs have had a high level of success, yet in other closely related areas such as computer vision and more complex recognition tasks there have been difficulties in obtaining a good level of performance. This suggests that there are areas which could be improved with potentially large benefits to technology. This thesis focuses mainly on HMM training, and uses a gesture recognition system application to verify the results of the synthetic data trials.
The training of HMMs is difficult because of overspecialization, local maxima traps, and the selection of the appropriate model structure and quantization scheme.
We focus our attention on two important methods: the Baum-Welch method and the Viterbi Path Counting method. We compare these two main methods with a range of other methods including Matthew Brand's Entropic MAP method, and Ensemble Averaging. We also investigate a range of different model structures to see which is the most important from the point of view of HMM training.
We produce a comparison of the performance of these methods on synthetic and real gesture data, and utilise model quality, classification performance, overspecialisation measures, and condition numbers to provide a picture of the learning process.
A set of rules are produced to guide the construction of HMM-based training systems for gesture recognition applications. These rules may also be useful in a wider range of applications.
Surprisingly good performance was obtained for the Viterbi Path Counting method. This was consistently the best method on all our trials. This method is very similar to one introduced by Stolcke and Omohundro, and also shares some similarities with the segmented k-means method of Rabiner, but there are several differences in implementation including the sequence order. We extended its scope by varying several key parameters including initial values and randomization to gain an improved understanding of its function.
In summary, the thesis investigates a range of methods to gain a better understanding of HMM-based gesture recognition and develops some key principles, testing the techniques on synthetic and real data.