A framework to detect and classify activity transitions in low-power applications

Minimizing the number of computations a low-power device makes is important to achieve long battery life. In this paper we present a framework for a low-power device to minimize the number of calculations needed to detect and classify simple activities of daily living such as sitting, standing, walking, reaching, and eating. This technique uses wavelet analysis as part of the feature set extracted from accelerometer data. A log-likelihood ratio test and hidden Markov models (HMM) are used to detect transitions and classify different activities. A tradeoff is made between power and accuracy.