Selection of Time-Domain Features for Fall Detection Based on Supervised Learning

Abstract

Latest mobile phones bring many advantages to our daily life. These gadgets are usually equipped with multiple sensors, such as camera, GPS, and accelerometer. The data collected from these sensors are used to develop several applications, such as face recognition, voice recognition, and daily activity analysis. Studies show that many accidents occur during daily activities, and we propose the idea that mobile phones can be used to distinguish these accidents from activities of daily living (ADL). Early intervention may reduce or eliminate the risk of fatal injury during an accident. In this study, we introduce a new mobile phone application designed for the recognition of falls and reporting the incident to appropriate authorities. We built a predictive model using supervised learning methods with selected features to detect falls with ratio as high as possible. 8 healthy people performed various activities carrying smart phone on their pockets to record simulated falls and ADLs to build our data set. The data-set consist of 43 time-domain features extracted from 3-axis accelerometer data. The analysis of several feature selection algorithms demonstrates that only 5 to 10 most discriminative features give the best success ratio for fall detection. Our test results show that the proposed system can recognize the fall events with approximately %90 success ratio.