Deterministic Initialization of Hidden Markov Models for Human Action Recognition

Human action recognition is often approached in terms of probabilistic models such as the hidden Markov model or other graphical models. When learning such models by way of Expectation-Maximisation algorithms, arbitrary choices must be made for their initial parameters. Often, solutions for the selection of the initial parameters are based on random functions. However, in this paper, we argue that deterministic alternatives are preferable, and propose various methods. Experiments on a video dataset prove that the deterministic initialization is capable of achieving an accuracy that is comparable to or above the average from random initializations and suffers from no deviation thanks to its deterministic nature. The methods proposed naturally extend to be used with other graphical models such as dynamic Bayesian networks and conditional random fields.