Sparsity-based Feature Extraction in Fall Detection with a Portable FMCW Radar

Abstract

Due to the aging population, fall detection is crucial for elderly health care and assisted living. Radar-based methods attract much attention for its potential for high accuracy, robustness, and privacy preservation. In this paper, sparsity-based feature extraction methods are proposed to extract robust time-Doppler features with physical meanings for the classification of fall and fall-similar motions. First, sparse representation theory is introduced and through Gabor-based sparse dictionary, sparse representation of the received signals can be achieved in time-Doppler domain. Then, corresponding sparse point maps consisting of a series of sparse solutions are obtained by OMP-based algorithm. Particularly, reconstructed signals can be utilized to demonstrate that sparse features preserve most information from original ones while ignoring noise interferences. Finally, experiments have been conducted to show its feasibility by achieving an average accuracy of 95% on fall detection.