A multi-sensor based pre-impact fall detection system with a hierarchical classifier

Fall is a major threat to elders' health. The goal of our study is to establish a pre-impact fall detection system to reduce the harm of falls. For the problem that single-sensor based system can't achieve high accuracy, we propose a multi-sensor based system, which can fuse the data from waist and thigh. Collected data are transferred to a computer or a cellphone using wireless Bluetooth technique. A discrimination analysis based pre-impact fall detection model is developed. Human activities can be classified into three categories (non-fall, backward fall and forward fall) using a hierarchical classifier. In order to improve the classification accuracy, optimal discriminant features are selected for each layer of classifier. Then, experiments are conducted and the results show that our method can both achieve high sensitivity and specificity as well as long lead time.