Shape boundary tracking with Hidden Markov Models

This paper considers a Hidden Markov Model (HMM) for shape boundary generating which can be trained to be consistent with human expert performance on such tasks. That is, shapes are defined by sequences of "shape states" each of which has a probability distribution of expected image features (feature "symbols"). The tracking procedure uses a generalization of the Viterbi method by replacing its "best-first" search by "beam-search" so allowing the procedure to consider less likely features as well in the search for optimal state sequences. Results point to the benefits of such systems as an aide for experts in depiction shape boundaries as is required, for example, in Cartography.