Context-aware fall detection using a Bayesian network

Abstract

Human activity recognition is regarded as one of the most important topics in ubiquitous computing. In this paper, we focus on recognizing falls. Falls are a leading cause of death among elderly people. Most existing fall detection techniques focus on studying isolated fall motion under restricted, clearly defined conditions, and thus suffer from a relatively high false positive rate induced by many other activities that resemble a fall. In this paper, we present an integrated fall detection framework that incorporates isolated fall detection algorithms with context information using a Bayesian network. The context information can include a person's age, personal health history, physiological measurements (such as respiration, blood pressure, heart rate, etc.), physical activity level and location. These additional sources of information are complement inputs to our framework to improve decision accuracy in recognizing activities such as a fall. A Bayesian network is constructed to structure the probabilistic dependencies between isolated fall detection result and various contextual sensor readings, and perform inference on the likelihood of a fall in a given context. Preliminary experimental results demonstrate that context information can play a significant role in improving fall detection accuracy and reducing both false negative and false positive rates. We also demonstrate that our probabilistic Bayesian model can produce informative inference results even when partial contextual information is observed.