It was predominantly used in speech recognition in the past and still being used where it deals with a one-dimensional signal. Now a number of articles mention its use for gesture recognition as well. Within a span of few months newer articles having solutions in HMM use to appear. Interestingly the problem still remain unsolved and still new papers on HMM queue up. How do they get aligned in timeline in a single thread of execution?
Is probabilistic estimation of a gesture (say, a trajectory for a two dimensional shape) would ultimately give a deterministic solution through abstract features?
I would appreciate explanation and justification from fellow researchers.
What you said about one dimensional signal is not entirely true. In speech recognition for example, each segment of a recording is described using sets of features like MFCC or PLP, which may amount to as much as about 40 features (depending on implementation). The HMM is used to model the dynamics od the input signal - each hidden state in HMM is trained to have different observation probability distribution, and e.g. for a 3-state HMM you can try to interpret the states as representing the probability of beginning, middle and ending of a vowel (or more generically: phone). I imagine that modelling such dynamics might be also beneficial for other classification problems you mentioned, like gestures, which might be modeled using a number of states adequate to their complexity.
A hidden Markov model is a Markov model in which the underlying state of the model is not uniquely recoverable from the observation stream. It is sometimes possible to recover the most probable underlying state timing, but the probabilities depend on the model itself, and even given a particular probability distribution assignment to transitions and outputs, the "true" timing of the underlying states is not necessarily recoverable.
From my weak knowledge about the topic, the states are what's hidden in HMMs. HMMs can be seen as a machine sealed in a black box that produces output based on 3 (or 4, or 5, or whatever) interconnected modules. One cannot know what's going on within the machine (what module is currently outputting data) but by studying the output one may figure out that the machine has 3 types of behavior. I don't know much about the mathematical side of the concept but I do know that much.
In a standard Markov model the states and transition probabilities between them are known, and one can generate sequences given the initial state. Given a particular sequence one can calculate the probability of it having been generated by the particular Markov model. In a hidden Markov model one still assumes an underlyig Markov model, but one can not directly observe the states, rather the states are related to the observables by emission probabilities. Given a sequence of observables, one can calculate the probability of it having been generated by a particulart HMM.
HMM was indeed originally developed for speech recognition, but can be used for modeling any stochastic temporal process, such as handwriting, musical scores, dance steps, recognizing a sequence of gestures, etc. These applications are not new ("the last few months"), but rather have been explored for many years now. The application of the technique to new domains is straightforward, so there is nothing really "open" - the only question is whether an appropriate HMM (and there are many variants!) is better or worse on a given data set than other techniques.
Y(J)S
- Badges
- Science topic
- Similar topics
- Medicine
- Physiology
- Biosignals
- Biomedical Signal Processing
- Statistical Analysis