Novel Composite Motion Extraction from Velocity Signature of FMCW Radar for Activity Recognition

Human activity recognition (HAR) is indispensable for in-home elderly assisted living and telehealth. We propose a novel technique of unobtrusive HAR by utilising a composite motion pattern (CMP) extraction procedure from a frequency modulated continuous wave radar. The technique considers the peak Doppler shift and the secondary Doppler shifts that are centered around it in the radar spectrum to effectively compute a velocity distribution for a single radar chirp. Multiple such distributions, when stacked together produce a CMP that is visibly distinct for each activity. Extracted features from these CMPs are classified and tested on external cohort to obtain an accuracy of 90.83%(±5.2) across all activities.