Interpretation of divers' symbolic language by using hidden Markov models

This paper reports the results of a system based on hidden Markov models (HMM) that is used to interpret both static and dynamic divers' hand signals using real time video feed. Two methods of collecting features that describe diver gestures are described and two types of HMMs are investigated: one based on discrete outputs variable distribution and the other based on mixture of Gaussians outputs variable distribution. For 8 basic diver gestures, 800 data samples were collected and the results were analyzed for the purpose of determining the most appropriate i) feature vector describing diver gestures, ii) HMM parameters (number of states and mixture components), and iii) type of the hidden Markov model. Finally, the quality of performance for recognition of each gesture is evaluated and analyzed, and necessary steps for improving the overall system are reported.