Automatic segmentation of piecewise constant signal by hidden Markov models

We propose an automatic signal segmentation algorithm for piecewise constant signals, which is based on Hidden Markov Models (HMM). It segments the observed data without the need for training data and initial conditions. One of the problems for automatic segmentation using HMM models is the determination of their number of states. In this paper, the number of states is estimated according to a maximum a posteriori (MAP) criterion. The proposed algorithm is iterative. Its initial conditions are chosen by a Tree-Structure technique, which is completely data driven. The segmentation is further improved by the multiscale technique. The performance is evaluated by computer simulations.