Near-Optimal Activity Prediction through Efficient Wavelet Modulus Maxima Partitioning and Conditional Random Fields

The quick evolution of sensors and the broad scale utilization of pervasive devices have awashed ubiquitous systems with an unprecedented amount of sensor data. Inferring activity or context from sensor data has fueled enormous research interests. In this paper, we propose a novel predictive model that utilizes wavelets, voronoi regions and Conditional Random Fields (CRF) to predict activities from accelerometer sensor data. In particular, our approach employs wavelet transform to decompose time-domain accelerometer sensor signals and extracts vital feature vectors from the resulting spectral. We introduce a new optimization technique to design a codebook during vector quantization of wavelet feature vectors. We couple the optimized codebook with CRF to craft a robust predictive model for activity recognition. To demonstrate the efficiency and effectiveness of our predictive model, we perform numerous experiments using accelerometer sensor data that emanates from android smart phones. Our technique yields a high overall prediction performance of up to 96.43%.