Recurrent Neural Networks Architectures for Accidental Fall Detection on Wearable Embedded Devices

Abstract

Unintentional falls can cause severe injuries and even death, especially if no immediate assistance is given. The aim of Fall Detection Systems (FDSs) is to detect an occurring fall in real time therefore issuing a remote notification. An accurate FDS can drastically improve the quality of life of elderly subjects or any other person at risk. In this chapter we focus on real-time Automatic Fall Detection (AFD) performed onboard smart wearable devices. In particular, in this chapter we discuss the feasibility of AFDs methods based on Deep Learning (DL) techniques that could fit the limited computation power and memory of smaller low-power Micro-Controller Units (MCUs). The chapter proves that a relatively simple Recurrent Neural Network (RNN) architecture, based on two Long Short-Term Memory (LSTM) cells, could be a viable candidate for embedded AFD. Tests were performed using the SisFall dataset, which includes sequences of tri-axial accelerometer and gyroscope readings for simulated falls performed by volunteers. This dataset was further annotated for training the RNN architecture. The resulting AFD method is shown to outperform other methods based on statistical indicators, as reported in the literature. The embedded feasibility of such approach is validated with an implementation for the SensorTile ^® by STMicroelectronics hardware architecture.