Accelerometer-based elderly fall detection system using edge artificial intelligence architecture

• Published in: International journal of electrical and computer engineering (Malacca, Malacca) Vol. 12; no. 4; p. 4430
• Main Authors: Zaid Salah, Osama, Selvaperumal, Sathish Kumar, Abdulla, Raed
• Format: Journal Article
• Language: English
• Published: Yogyakarta IAES Institute of Advanced Engineering and Science 01.08.2022
• Subjects: Accelerometers; Algorithms; Artificial intelligence; Artificial neural networks; Computer architecture; Deep learning; Fall detection; Internet of Things; Machine learning; Microcontrollers; Model accuracy; Network latency; Neural networks; Older people; Power consumption; Real time; Wide area networks
• ISSN: 2088-8708; 2722-2578; 2088-8708
• DOI: 10.11591/ijece.v12i4.pp4430-4438

Falls have long been one of the most serious threats to elderly people's health. Detecting falls in real-time can reduce the time the elderly remains on the floor after a fall, hence avoiding fall-related medical conditions. Recently, the fall detection problem has been extensively researched. However, the fall detection systems that use a traditional internet of things (IoT) architecture have some limitations such as latency, high power consumption, and poor performance in areas with unstable internet. This paper intends to show the efficacy of detecting falls in a resource-constrained microcontroller at the edge of the network using a wearable accelerometer. Since the hardware resources of microcontrollers are limited, a lightweight fall detection deep learning model was developed to be deployed on a microcontroller with only a few kilobytes of memory. The microcontroller was installed in a low-power wide-area network based on long range (LoRa) communication technology. Through comparative testing of different lightweight neural networks and traditional machine learning algorithms, the convolutional neural network (CNN) has been shown to be the most suited, with 95.55% accuracy. The CNN model reached inference times lower than 37.84 ms with 61.084 kilobytes storage requirements, which implies the capability to detect fall event in real-time in low-power microcontrollers.