Optimal Number of States in Hidden Markov Models and its Application to the Detection of Human Movement
In this paper, Hidden Markov Model is applied to model human movements as to facilitate an automatic detection of the same. A number of activities were simulated with the help of two persons. The four movements considered are walking, sitting down-getting up, fall while walking and fall while standing. The data is acquired using a biaxial accelerometer attached to the person’s body. Data of the four body gestures were then trained to construct several Hidden Markov models for the two people. The problem is to get a good representation of the data in terms of the number of states of the HMM. Standard general methods used for training pose some drawbacks i.e. the computational burden and initialisation process for the model estimate. For this reason, a sequential pruning strategy is implemented to address the problems mentioned.
Keywords: Hidden Markov Models, sequential pruning strategy, Bayesian Inference Criterion