Covariance matrix based fall detection from multiple wearable sensors

Falls are among the critical accidents experienced by elderly people and patients carrying some diseases. Subsequently, the detection and prevention of falls have become a hot research and industrial topic. This is due to the fact that falls are behind numerous irreversible injuries, or even death, and are weighting on the budgets of the health services. Automatic fall detection is one of the proposed solutions which aim at monitoring people who are likely to fall. Such solutions mitigate the fall impact by taking a quick action, e.g. in case of a fall occurrence, an alert is sent to the hospital. In this paper, we propose a new fall detection system relying on different signals acquired with multiple wearable sensors. Our system makes use of the covariance of the raw signals and the nearest neighbor classifier. Besides feature extraction, we also employ the covariance matrix as a straightforward mean for fusing signals from multiple sensors, to enhance the classification performance. Evaluation on two publicly available fall datasets, namely CogentLabs and DLR, demonstrates that the proposed approach is efficient when exploiting a single sensor as well as when fusing data from multiple sensors. Geodesic metrics are found to provide a higher fall detection accuracy than the Euclidean metric. The best obtained classification accuracies are 92.51% and 98.31% for CogentLabs and DLR datasets, respectively.