Catching Attention Movements: Real-Time Camera-Based Anomaly Detection for Patient Care

Abstract

In this study, we introduce Catching Attention Movements (CAM), defined as brief, nonperiodic motions that deviate from typical scene dynamics and naturally draw attention. Detecting CAM is clinically valuable for monitoring patients across diverse conditions and for improving vital-sign measurements by identifying or excluding motion-affected segments. We present a real-time CAM detection and alerting method that uses a standard USB camera as a remote optical sensor. Motion dynamics are reconstructed via optical flow, and clustering across time, position, and scene identifies candidate events. An adaptive threshold then determines whether a candidate qualifies as a CAM, triggering an alert when appropriate. We evaluate the method in two real-world settings: (A) seizure-onset registration in epilepsy patients using 50 recordings of major motor seizures, and (B) infant vital-sign monitoring using a 6-hour video dataset. Results show reliable seizure detection and improved confidence in physiological measurements, underscoring CAM detection’s clinical utility.