Toward Accurate Fall Detection with a Combined Use of Wearable and Ambient Sensors

Abstract

This paper proposes the use of relative barometric pressure sensors for enhancing the accuracy of fall detection based on a tri-axial accelerometer. In this study, a device consisting of a tri-axial accelerometer and a barometric pressure sensor was placed on a subject's waist and another reference barometric pressure sensor is placed on the wall. The data is collected while the subject is performing a sequence of 22 fall and non-fall activities 12 times. We compare activity classification with features extracted from these following datasets: 1) acceleration data, 2) acceleration and on-body barometric pressure, and 3) acceleration, on-body barometric pressure, and the reference barometric pressure. Three classifiers are compared, i.e., k-Nearest Neighbors, Decision Tree, and Random Forest. The results show that Random Forest yields the highest classification accuracy among the three classification algorithms and accuracy enhancement can be achieved with the relative barometric pressure information. The highest classification accuracy of 96.6% can be achieved with the Random Forest algorithm when acceleration, on-body barometric pressure, and the reference barometric pressure information on the wall are used.