Detection of typical abnormal behavior in home-based elderly care based on ViT-iECGAN significant information migration compensation

In the context of solitary living, the real-time surveillance of abnormal behaviors is of paramount importance for ensuring the well-being and safety of the elderly. A prevalent concern is the occurrence of unexplained falls, which may stem from underlying pain or illness. Furthermore, abusive behaviors within home care settings are a significant issue. It is evident that failure to detect these behaviors promptly can result in severe health complications, potentially culminating in fatal outcomes. However, the existing video surveillance systems have inherent privacy protection flaws, which impede their widespread adoption for in-home behavioral monitoring of the elderly. In light of the recent surge in video privacy breaches, there is an urgent need for a computer vision-based anomalous behavior detection method that upholds stringent video privacy standards. This paper introduces an innovative approach that integrates compressed sensing for visual privacy preservation, along with significant information migration compensation and behavior detection utilizing generative adversarial networks (GAN). There are three main steps: (1) multilayer compressed sensing using stage transformation to obtain visual privacy preserving state video; (2) combining the attention mechanism of ViT with iECGAN to form a ViT-iECGAN model that labels salient regions of the original video; and (3) ViT-iECGAN model migration is used to compensate for salient information with detection of behaviors. Experimental results on some publicly available behavioral datasets show that the method has better prediction and good generalization performance.