Smart Vision-Based Action Monitoring in Assistive Robots for Elderly Safety

Abstract

This paper presents a comprehensive review of vision-based action recognition models for use in medical assistive robots, with a focus on improving indoor safety for the elderly. We evaluate various architectures, including 3D convolutional neural networks, graph convolutional networks, transformer-based models, and pose-based detectors, based on criteria such as recognition accuracy, inference latency, and computational efficiency. Real-world robotic implementations are also compared to highlight current capabilities. Based on the analysis, lightweight GCNs and compact pose estimation models are identified as strong candidates for deployment on resourceconstrained platforms, such as the Kompai robot. In addition, a practical framework is proposed to guide model selection and system integration for developers and researchers in the field.