Multivariate Hidden Markov Models for Personal Smartphone Sensor Data: Time Series Analysis

Abstract

Smartphone-based human activity recognition (HAR) offers growing value for health research. We applied offline Hidden Markov Models (HMMs) to multivariate smartphone sensor data, classifying individual behaviour into a time series of states. We used supervised HMMs, validated using ground-truth data from a small self-report study. The HMMs achieved reasonable accuracy in classifying phone off-person vs. phone on-person, off-vehicle vs. on-vehicle, and phone off-person vs. sitting vs. standing vs. walking, for some participants. Strong evidence suggests that poor accuracy in other cases was caused by participant mislabeling, though HMM shortcomings contributed.