A semi-supervised Hidden Markov model-based activity monitoring system

Most existing human activity classification systems require a large training dataset to construct statistical models for each activity of interest. This may be impractical in many cases. In this paper, we proposed a semi-supervised HMM based activity monitoring system, that adapts the HMM for a specific subject from a general model in order to alleviate the requirement of a large training data set. In addition, using two triaxial accelerometers, our system not only identifies simple events such as sitting, standing and walking, but also recognizes the behavior or a more complex activity by temporally linking the events together. Experimental results demonstrate the feasibility of our proposed system.