Optimal Sensor Placement for Human Activity Recognition with a Minimal Smartphone-IMU Setup

Abstract

Human Activity Recognition (HAR) of everyday activities using smartphones has been intensively researched over the past years. Despite the high detection performance, smartphones can not continuously provide reliable information about the currently conducted activity as their placement at the subject’s body is uncertain. In this study, a system is developed that enables real-time collection of data from various Bluetooth inertial measurement units (IMUs) in addition to the smartphone. The contribution of this work is an extensive overview of related work in this field and the identification of unobtrusive, minimal combinations of IMUs with the smartphone that achieve high recognition performance. Eighteen young subjects with unrestricted mobility were recorded conducting seven daily-life activities with a smartphone in the pocket and five IMUs at different body positions. With a Convolutional Neural Network (CNN) for activity recognition, activity classification accuracy increased by up to 23% with one IMU additional to the smartphone. An overall prediction rate of 97% was reached with a smartphone in the pocket and an IMU at the ankle. This study demonstrated the potential that an additional IMU can improve the accuracy of smartphone-based HAR on daily-life activities.