Fall Detection from UWB Radars: A Comparative Analysis of Deep Learning and Classical Machine Learning Techniques

Abstract

Fall detection poses significant challenges for researchers due to the potential for severe injuries, such as femoral neck fractures, brain hemorrhages, and skin burns, which can cause considerable pain. Furthermore, undetected falls can lead to deteriorating health over time, resulting in a painful end-of-life scenario or even death. It is vital to efficiently detect falls in order to promptly notify relevant individuals, such as nurses, for timely intervention. This study presents a solution for fall detection specifically designed for healthcare institutions, with a focus on individuals aged 65 and over. Ultra-wideband (UWB) radars are employed as the primary technology for this purpose. Two distinct approaches were implemented: one involving feature extraction, dimensionality reduction, and a Random Forest classifier, and the other utilizing a simple Convolutional Neural Network (CNN) architecture. The results consistently demonstrate that the deep learning approach outperforms the classical machine learning approach by an average margin of 7.5% and 8.1% when all raw data are filtered either by Butterworth or Type 1 Chebychev filters, respectively. The superior performance of the deep learning approach can be attributed to the effective filtering process applied to the UWB radar data. Importantly, a leave-one-subject-out strategy was employed to validate the fall detection performance. This strategy is not commonly used for validation in other research studies present in the literature.