Smart Care System with Dual-Model Cooperative Triggering for Anomaly Detection

This paper introduces an intelligent monitoring system designed for the timely detection of abnormal postures in elderly residents within care facilities. The system integrates deep learning-based posture detection, utilizing YOLOv10n, with facial recognition, employing a synergy of YOLOv10n-face and VGGFace2, to ensure accurate resident identification. A novel event-triggering mechanism, predicated on dual bounding box overlap, activates when the abnormal posture bounding box fully encapsulates the facial bounding box, achieving an Intersection over Union (IoU) of 1. Upon triggering, the system logs the event and its duration, storing the data in a backend database. If the abnormal posture persists beyond a pre-defined temporal threshold, immediate alerts are dispatched to caregivers via the LINE API. Authorized personnel can access real-time RTSP video streams and anomaly logs of assigned residents through a role-based access control (RBAC) web platform. The proposed system demonstrates high accuracy and responsiveness, indicating its significant potential for practical deployment in smart healthcare environments.