Radar-Based Action Classification for Elderly Healthcare in Caregiving Environment

Monitoring multiple actions of the elderly, including lying, sitting, standing, bed-exit, and falling, is crucial in caregiving environments to prevent adverse events and reduce caregiver workload. Unlike many existing studies that focus solely on fall detection or use simulated environments, this study collects data on these five actions using a real electric nursing bed, providing a more realistic representation of practical care scenarios. A millimeter-wave radar–based dataset is developed with recordings from 15 participants, and a comparative evaluation is conducted across multiple deep learning classification networks. The performance of models such as ResNet18, ResNeXt50, MobileNetV3, ShuffleNetV2, and Vision Transformer is assessed in terms of both accuracy and model size. Experimental results show that MobileNetV3-Large achieves the highest accuracy of 80.49%, while ShuffleNetV2 attains the smallest model size of 5.45 MB, making it particularly suitable for real-time deployment on edge devices. These findings provide practical guidance for selecting lightweight yet accurate models for elderly care monitoring systems.