A comparison of automatic fall detection by the cross-product and magnitude of tri-axial acceleration

Abstract

Falling is an important problem in the health maintenance of people above middle age. Portable accelerometer systems have been designed to detect falls. However, false alarms induced by some dynamic motions, such as walking and jumping, are difficult to avoid. Acceleration cross-product (AC)-related methods are proposed and examined by this study to seek solutions for detecting falls with less motion-evoked false alarms. A set of tri-axial acceleration data is collected during simulated falls, posture transfers and dynamic activities by wireless sensors for making methodological comparisons. The performance of fall detection is evaluated in aspects of parameter comparison, threshold selection, sensor placement and post-fall posture (PP) recruitment. By parameter comparison, AC leads to a larger area under the receiver operating characteristic (ROC) curve than acceleration magnitude (AM). Three strategies of threshold selection, for 100% sensitivity (Sen100), for 100% specificity (Spe100) and for the best sum (BS) of sensitivity and specificity, are evaluated. Selecting a threshold based on Sen100 and BS leads to more practicable results. Simultaneous data recording from sensors in the chest and waist is performed. Fall detection based on the data from the chest shows better global accuracy. PP recruitment leads to lower false alarm ratios (FR) for both AC- and AM-based methods.