Activity recognition from sensors using dyadic wavelets and Hidden Markov Model

Roland Assam, Thomas Seidl

Abstract: Advances in sensor and ubiquitous technologies have contributed to the broad scale adoption of pervasive devices. Context or activity recognition from sensor signals is an emerging area that has garnered huge research interest. In this paper, we propose a novel predictive model that utilizes dyadic wavelet transform, vector quantization and Hidden Markov Model (HMM) to predict a high level activity from low level accelerometer sensor signals. Specifically, we analyze and extract important spectral features of the sensor signal by performing multi-resolution wavelet transform. These features are utilized to institute a codebook through the process of vector quantization. An enhance HMM predictive model for activity recognition is built using the codebook and some wavelet feature vectors. We conducted numerous experiments using accelerometer sensor data stemming from android smart phones. Our experiments reveal superior prediction results with a prediction accuracy of up to 96.15%.