Online Fall Detection Using Attended Memory Reference Network

Falls cause serious injuries that make daily activities difficult; therefore, they are a common target action for intelligent monitoring systems. Existing vision-based methods for fall actions classify well-trimmed short videos as either fall or non-fall actions. However, critical limitations exist when applying these methods to untrimmed videos including fall and non-fall actions as well as background. These methods can determine whether there is a fall or not for an input video with many frames related to either fall or non-fall actions. In addition, these methods require offline processing for a whole video as input, while there is strong demand for quicker responses to fall injuries provided by online fall detection. To this end, we introduce an attended memory reference network that detects a current action online for a given video segment consisting of past and current frames. To integrate contextual information used for detecting a current action, we propose a new recurrent unit, called an attended memory reference unit, which accumulates input information based on visual memory attended by current information. In an experiment using a fall detection dataset obtained from the abnormal event detection dataset for CCTV videos publicized by AI Hub, the proposed method outperforms state-of-the-art online action detection methods. By conducting ablation studies, we also demonstrate the effectiveness of the proposed modules related to the attended visual memory.