Fall Detection with Privacy as Standard

Current ambient assisted-living (AAL) systems used to detect falls in the elderly often rely on the person in question being either capable of pressing a panic button to alert others of their fall, or a wearable device which detects impacts. These systems can be invasive, negatively impacting on an individual's sense of independence and privacy which may in turn lead to a lower quality of life. This paper aims to examine a non-invasive means of detecting falls within the home, focusing on three separate approaches to detection, a waist-worn, computer-vision-based and novel in-shoe systems. The effectiveness of each individual system is explored and their effectiveness in detecting falls versus the privacy they provide are examined. The machine learning model which was trained for use with the waist-worn system achieved an accuracy of 74%, with a sensitivity of ~ 70%, using the limited dataset available for this preliminary study. The computer-vision system can accurately detect individuals in a scene as well as fall scenarios, however drastic lighting changes negatively impact the systems performance. Our in-shoe system achieved a zero false positive rate with an accuracy of ~ 67%.