Fall Detection Approach Using Variational Autoencoders with Self-Attention Features

Abstract

In this paper, we propose an alternative method for fall detection using variational autoencoders (VAEs) with an attention mechanism on an existing dataset. The dataset consists of 6 different fall cases from 21 people. For effective fall detection, we introduce the use of the magnitude of the acceleration vector (MAV) of wearable gyroscope data and apply fast-Fourier transform (FFT) to create new features. These FFT features are then passed through attention modules with self-combination to form attention features. Our experimental results show that the VAE with self-attention features achieved an average accuracy of 90.7% and an F1 score of 93.8% in fall detection, demonstrating the effectiveness of the proposed method in utilizing gyroscope sensors for fall detection in the context of threshold criteria.