PSD-RNN for Radar-Based Elderly Fall Detection: Integration of Power Spectral Density and Recurrent Neural Networks

Abstract

The rapidly growing, aging, and increasingly isolated structure of society highlights the importance of elderly care and emergency detection systems. Falls are among the most prevalent medical emergencies, which the vast majority of individuals will, at some point, be confronted. In this study, radar-based fall detection is performed using the Power Spectral Density (PSD)-Recurrent Neural Network (RNN) method. Unlike computationally heavy image-based CNN approaches dominant in the literature, utilizing 1D PSD features significantly reduces input dimensionality while effectively capturing frequency-domain temporal dynamics. The PSDs of a total of 700 fall and daily activity data measured by continuous-wave radar are calculated and used as features. Subsequently, a grid search is performed, and classification is carried out as a result of a broad RNN scan with 3240 hyperparameters. The method distinguishes between falls and non-falls with 98.38% accuracy in various scenarios using a subject-based Leave One Subject Out (LOSO) cross-validation. This success also demonstrates its effectiveness in activity recognition, achieving 91.27% validation accuracy and 95.24% unseen subject test accuracy. The performance of the method is also presented by means of a comparison with various other methods. It is anticipated that the proposed method will reach a wider range of applications with different solution-oriented strategies and future optimizations.