Improving the accuracy of a fall detection algorithm using free fall characteristics

The number of elderly who live on their own is increasing. Falls in the elderly are a major problem which has a long term effect. If the elderly can get help immediately after a fall, the severity of the injury, and the cost could be reduced. This paper presents a fall detection algorithm utilizing two thresholds for the resultant acceleration in 1.5-s window segments. Data was acquired using a tri-axial accelerometer mounted on the user's trunk, and two types of experiments were investigated. In Type A), ten young subjects performed falls and typical Activity of Daily Living (ADL) tasks. In Type B), ten young subjects performed falls, where ten elderly subjects performed ADL tasks. There were four categories of fall: forward fall, backward fall, left and right side fall, and six categories of ADL: sit-stand, stand-sit, sit-lie, lie-sit, bend down, and walking 2 m. The results, tested on 300 sequences for each type, show that falls can be distinguished from ADL with 100% sensitivity and 93.33% specificity for Type A, and 100% sensitivity and specificity for Type B. Improval performance of the algorithm was possible by comparing with a threshold obtained from the maximum peak resultant acceleration. Moreover, this algorithm has low computational complexity, which allows for its real-time implementation on a microcontroller.