Compressed Neural Network for Thermal Array-Based Fall Detection System on Embedded AI

Fall incidents may lead to more serious health issues if not promptly treated. Existing fall detection systems mostly use cameras and are considered a privacy-intrusive approach. Thermal array sensors are considered a privacy-friendly device that does not raise discomfort for users. In this study, we presented a fall detection system using a thermal array sensor with three different deep learning approaches: LSTM, BiLSTM, and GRU. Our model is optimized using the pruning method to further efficiently deployed into Embedded AI. Our result shows that BiLSTM has the most promising result by 99.93% accuracy, 99.73% precision, and 0.057% False Alarm Rate (FAR).