Peningkatan Akurasi Deteksi Jatuh Menggunakan Sensor Akselerometer dan Giroskop pada Smartphone

Abstract

The aging population is a global concern, partly because as the body ages, physical conditions weaken, increasing the likelihood of falls. Falls are particularly dangerous for the elderly as they can lead to serious problems and even death. Detecting falls quickly and accurately is crucial to implement preventive measures and timely intervention when a fall occurs.

This research focuses on designing a human physical activity classification system, primarily used for fall detection. Seven model architectures are proposed using a novel approach involving the variant of recurrent neural network (RNN) methods, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Neural Network (SimpleRNN). Additionally, variations with Convolutional Neural Network (CNN) are explored, specifically 1D Convolutional Neural Network (1D CNN).

Validation results of the classification show that the experimented methods for the classes of sitting, standing, and falling achieved perfect scores, while the falling class showed varying scores for each designed model architecture. For the overall classes, the lowest performance is observed in the combination of 1D CNN and SimpleRNN architecture with an accuracy of 95.6%, whereas the highest performance is attributed to the SimpleRNN architecture and the combined CNN and GRU architecture with an accuracy reaching 99.0%.