I have times series data from accelerometer that was attached to a person that was doing different type of exercises. I have a feature matrix that is basically a table with 3 columns (3-axis acceleration). I also have a groundtruth verctor with labels corresponding to the exercises. Now I want to train a Hidden Markov Model (HMM) using this feature matrix and the vector with labels. For that I am splitting my feature matrix into chunks where in each chunk I have only the samples of a particular exercise. Then I am training my HMM.
Now, on the testing phase I have a feature matrix without groundtruth. How should I feed my samples to the HMM? I can't create chunks anymore since I do not have groundtruth.
In general, is my approach valid? I am trying to think the same way as in word recognition scenario. There we have chunks with samples and each chunk is one word...
