Patimakorn Jantaraprim, Pornchai Phukpattaranont, Chusak Limsakul, and Booncharoen Wongkittisuksa

pp. 637 - 644

Abstract

We present a system for improving fall detection performance using a short time min-max feature based on the specific
signatures of critical phase fall signal and a neural network as a classifier. Two subject groups were tested: Group A involving
falls and activities by young subjects; Group B testing falls by young and activities by elderly subjects. The performance was
evaluated by comparing the short time min-max with a maximum peak feature using a feed-forward backpropagation network
with two-fold cross validation. The results, obtained from 672 sequences, show that the proposed method offers a better
performance for both subject groups. Group B’s performance is higher than Group A’s. The best performances are 98.2%
sensitivity and 99.3% specificity for Group A, and 99.4% sensitivity and 100% specificity for Group B. The proposed system
uses one sensor for a body’s position, without a fixed threshold for 100% sensitivity or specificity and without additional
processing of posture after a fall.