Threshold-Based Algorithm for Fall Detection Through an Inertial Sensor Fixed at the Head

Abstract

This work was focused on developing a low-cost, effective and reliable algorithm to be applied on fall detection through the analysis of data provided by an Inertial Measurement Unit attached at the head. In order to achieve this goal, trends of the acceleration vector variance were analyzed for the case of fall and no fall conditions, and such a feature, together with the time at which this feature remained above certain threshold, were the key points used for assessing whether a fall had happened or not. A public dataset, which included different types of falls and daily activities, was used for algorithm validation. The algorithm parameters (thresholds) were adjusted to achieve the highest Recall value. Accordingly, values of Precision, Recall, Specificity and Accuracy equal to 92.8%, 98.7%, 90.4%, and 95.0%, respectively, were obtained.