Histogram-Based Training Initialisation of Hidden Markov Models for Human Action Recognition

Human action recognition is often addressed by use oflatent-state models such as the hidden Markov model andsimilar graphical models. As such models requireExpectation-Maximisation training, arbitrary choicesmust be made for training initialisation, with major impacton the final recognition accuracy. In this paper, wepropose a histogram-based deterministic initialisation andcompare it with both random and a time-baseddeterministic initialisations. Experiments on a humanaction dataset show that the accuracy of the proposedmethod proved higher than that of the other testedmethods.