Fall Detection System Using Accelerometer and Gyroscope Sensors with Backpropagation Neural Networks

Abstract

A fall is an incident in which a person suddenly loses their balance and falls uncontrollably to the ground or another surface. This is particularly dangerous for the elderly, as it can lead to serious injury or even death if not treated immediately. This study suggests utilizing a Backpropagation Neural Network (BPNN) to classify fall and non-fall incidents, particularly in older adults, by measuring alterations in linear acceleration and angular velocity obtained from accelerometer and gyroscope sensors. A dataset featuring eight distinct movement types, including standing to bowing, standing to squatting, sitting to standing, standing to prostrating, sitting to prostrating, falling sideways, falling forward, and falling backward, is collected from 69 participants and processed by determining the difference (Δ) between consecutive sensor measurements, followed by normalization, labeling, and statistical feature extraction. The collected features are fed into a BPNN that has two hidden layers, each with 64 and 128 neurons. The proposed model achieved an accuracy of 86.87%, validation accuracy of 85.00%, and testing accuracy of 92.86%. Furthermore, test results show a precision of 94%, recall of 85%, and specificity of 97%, demonstrating the system's ability to accurately classify eight different human activities and distinguish falls from normal daily activities. These findings indicate that a Backpropagation Artificial Neural Network in conjunction with the suggested Δ-based preprocessing offers an efficient method for inertial sensor-based fall detection systems.