Grammar-based, posture- and context-cognitive detection for falls with different activity levels

Falls are dangerous for the aged population as they result in serious detrimental consequences. Therefore, many fall detection methods have been proposed. Most of these methods characterize falls by large accelerations and fast body orientation changes. However, certain activities like sitting down quickly, vigorous gaits, and jumping, also show these characteristics, and thus are hard to distinguish from real falls. Moreover, many falls in the elderly are slow falls which show lower activity levels. Existing work fails to detect slow falls effectively because they only identify falls with high activity levels. In this paper, we present a grammar-based fall detection framework which not only better distinguishes fall-like activities from real falls, but also emphasizes the detection of slow falls. We utilize posture information extracted from on-body sensors and context information collected from sensors deployed in the house to reduce false positives. A fall in our framework is detected as a sequence of sensor events. We provide a context-free grammar to define these sequences so that the framework can be easily extended to detect more kinds of falls. Our case study shows that our method can distinguish various fall-like activities from real falls and can also effectively detect both fast falls and slow falls. The integration evaluation shows that our method achieves both high sensitivity and high specificity.