Real-time, model based algorithm implementation for human posture classification

A generic platform for continuously and unobtrusively monitoring human motion activity is deployed. Wirelessly transmitted data from a single three-axis accelerometer integrated into the headband is collected in real time on a laptop, and then analyzed to extract two sets of features necessary for posture/movement classification. The received acceleration signals is decomposed with discrete wavelet transform (DWT) to extract the first set of features; any change of the smoothness of the signal that reflects a transition between postures is detected at the finer DWT resolution levels. Fuzzy logic inference system (FIS) then uses the previous posture transition and the second set of features to choose one of eight different posture categories, namely sit, stand, lie on back, lie on left, lie on right, bend, walk, and run. Using the classifier in typical everyday activity among multiple users indicated more than 96.9%, 94.2%, 97.5% accuracy in detecting the static postures, walking, and running, respectively. Identifying the dynamic transitions among these steady postures achieved 92.6% accuracy.