Human Activity Recognition using Ceiling-Mounted Multi-Radar Sensors

Abstract

Continuous activity monitoring in assisted living should work without cameras or wearable devices. Radar meets this requirement, but most radar studies use a single sensor and activities recorded in isolation, and few deliver results in real time. We present an indoor monitoring system in which point clouds from three ceiling-mounted radar sensors are fused using density-based clustering, movement and posture features are derived for activity classification, and results are shown on a real-time dashboard. In a preliminary study with eight participants performing everyday activities as one continuous scenario, a random forest reached 74% frame-level accuracy with a random 80/20 split and 68% with leave-one-participant-out evaluation. Without position features, accuracy fell to 46% and 44%, indicating partial dependence on room layout, while window-level features without position reached 67% and 65%.