Introduction to hidden Markov models and their applications to classification problems

• Main Author: Zambartas, Michail
• Other Authors: Fargues, Monique P.
• Language: en_US
• Published: Monterey, California. Naval Postgraduate School 2012
• Online Access: http://hdl.handle.net/10945/8566

Approved for public release; distribution is unlimited === This thesis presents an introduction to Hidden Markov models (HMM) and their applications to classification problems. HMMs have been used extensively to model the temporal structure and variability of speech and other signals in the last decade. We selected to write our own HMM implementation in MATLAB. We tested our software on a limited isolated 4-word recognition. We also applied our implementation to the recognition of mine-like objects buried in shallow sand, using seismo-acoustic data obtained from an on-going project at the Naval Postgraduate School. Initial results indicate that the HMM-based classifier can recognize the type of mine-like object, independent of the object weight with a 97% accuracy. Results also indicate that it can recognize the object type at different distances with a 100% accuracy. However, the experiments were conducted with very few data, and further work needs to be done to confirm these initial findings by using a larger data set. Finally, we benchmarked our results against those obtained using a back-propagation neural network implementation, which were found to be similar, but slower than the HMM- based implementation.