A Multi-sensor Fall Detection System Based on Multivariate Statistical Process Analysis

Abstract

Physical injuries of elders resulting from fall accidents have become a serious social problem. An automatic fall detection system can provide timely feedback regarding the activities of elders and improve the quality of life of the elderly. This article aims to design an automatic fall detection system using multiple wearable sensors and find optimal sensor attachment locations. Six sensors are fixed on different parts of the human body to obtain real-time movement data. The autoregressive integrated moving average model is adopted to remove autocorrelation from the original data. Then, the multidimensional data are processed using principal component analysis. Finally, a novel threshold method based on a multivariate control chart is proposed for detection falls. This method can achieve high detection accuracy and can adapt to different individuals because the detection threshold is established using individual historical data. Different activities and simulated falls were performed by volunteers. Compared with most other single-sensor systems, the proposed multi-sensor system achieved higher sensitivity (94.8%) and specificity (95.2%), especially at a low sampling frequency (20 Hz). By evaluating the performance of different sensor attachment locations, the results showed that three sensors fixed on the waist, arm and thigh were able to efficiently distinguish falls from non-falls.