Project
Phase-1: Training the HMM model
HMMs are a well known concept used mainly in the field of speech recognition. In this project, I've used IMU data, taken from a phone, to classify gestures using HMM. The biggest problem with HMMs is that the state space increses exponentially with time. hence, it is unstable for a longer dataset. Hence I use something called the 'Forward-Backward' algorithm with the Expectation-Maximization technique to determine model parameters for various gesture models.
Phase-2: Testing: Estimating max log likelihood
Given the model parameters, I now compute the max log likelihood of the unknown dataset to classify it as one of the trained gestures. The figure here shows the probability of the unknown dataset belonging to either of the 6 gestures. Since the 'Infinity' label has the max likelihood, the unknown dataset is classified as 'Infinity' gesture.