Probabilistic Approach to Human Activity Recognition from Accelerometer Data

Abstract

Due to the wide availability of personal mobile and wearable devices, it is now easier than ever before to obtain information from a wide range of sources, including the actions performed by the users or the environment in which a user is located. This is advantageous in a number of applications including, in particular, remote health monitoring and diagnostics, where the manner in which an action is performed or the conformance to some treatment or living regime may be relevant to treatment or health outcomes. The goal of human activity recognition is to determine the identity of the action the user is currently performing based on the data streamed from some sensor modality. In this work we consider the accelerometer signals streamed through a wearable IMU unit and use this data to recognize the user's activity. We consider typical activities of daily living (ADLs). We adopt a simple approach based on the probabilistic modeling of the streamed signals. Some of the training samples are taken as templates and are modeled by empirical distributions. Testing for new sample consists essentially of measuring the probabilistic distribution of the test sample against those of the templates using a chosen set of distance/dissimilarity measures. This approach has proven to be both very computationally efficient and effective regarding its predictive performance.