Robust pre-impact fall detection algorithms and practical application in wearable airbag systems for injury prevention in older people

Abstract

Falls are a major public health problem due to their high prevalence and severe consequences in older population. Pre-impact fall detection, which aims to detect a fall accident before the body hits the ground, provides a radical solution to this problem because it allows some buffer time to deploy on-demand fall protection devices such as wearable airbags to directly prevent fall-related injuries. Existing algorithms suffered from poor robustness and rarely considered real applications. This dissertation composes of five consecutive studies which not only focus on the fundamental development of robust pre-impact fall detection algorithms, but also consider their practical applications to real-life wearable airbag systems in older people. The first study established a large-scale open motion dataset (KFall) using young participants based on wearable inertial sensors with synchronized high-frequency cameras. Based on the established dataset, the second study proposed a novel hybrid deep learning model (ConvLSTM) for pre-impact fall detection and comprehensively compared the accuracy and practicality of different types of state-of-the-art algorithms. The third study introduced a novel application of data augmentation to address various sensor rotation errors for pre-impact fall detection in practical wearable systems in which sensor displacements are inevitable due to loose attachment and body movements during long-time deployment. The fourth study validated the feasibility of applying algorithms trained by young subjects to older users under both laboratory and semi-naturalistic environment. The fifth study introduced a novel semi-supervised method for pre-impact fall detection based on limited fall data given practical difficulties in collecting real-world falls and time-consuming labeling process. The research outcome could inspire future research to shift their focus from traditional threshold-based or machine learning algorithms to cutting-edge deep learning algorithms. It can also facilitate algorithm development towards to a more practical direction which can be seamlessly embedded into real-life wearable systems for older people.