A wearable pre-impact fall detector using feature selection and Support Vector Machine

Falls and the resulting injuries in the elderly are a major public health problem, thus the early detection of falls is of great significance. The purpose of this study was to investigate the feasibility of a novel pre-impact fall detector prototype capable of detecting impending falls in their descending phase before the body hits the ground. A wearable tri-axial MEMS accelerometer was used for data collection of human motion information and a pair of wireless transceivers was used to transmit acceleration data to a PC for data analysis. Feature vector derived from time-domain characteristics was generated and feature selection was then performed to obtain the features with the most discrimination power. Fall detection algorithm using Support Vector Machine was developed and evaluated. The overall system was tested and results showed that all falls could be detected with an average lead-time of 203ms before impact, and no false alarm occurred. The proposed system will lead to potential applications for preventing or reducing fall-related injuries.