Next-generation fall detection: harnessing human pose estimation and transformer technology

Abstract

ABSTRACT Elderly falls are occurring at an alarming rate, with significant health risks for seniors. Current fall detection systems often lack accuracy, efficacy, and privacy considerations. This study examines three leading human pose estimation frameworks combined with transformer deep learning models to develop a lightweight, privacy-preserving fall detection system. Key features include: 1) It runs on low-power devices like Raspberry Pis; 2) It monitors seniors passively, without requiring active participation; 3) It can be deployed in any residential or senior care setting; 4) It does not rely on wearables; and 5) All processing occurs locally, ensuring privacy with only fall alerts transmitted to caregivers. In real-world tests, the model achieved 95.24% sensitivity, 89.80% specificity, 98.00% accuracy, a 90.91% F1 score, and 95.24% precision, highlighting its effectiveness in detecting falls among the elderly while maintaining privacy and security.