Action Recognition for Privacy-Preserving Ambient Assisted Living

The care challenges posed by an increasing elderly population have made ambient assisted living a significant research focus. Computer vision-based technologies can monitor older adults’ daily activities in their homes, providing insights into their health and prolonging their capacity to live independently. However, despite the benefits of these technologies, their widespread adoption has been hampered due to privacy concerns. These concerns frequently stem from the need to stream user data to cloud servers for computation, posing a risk to user privacy. This study proposes a privacy-preserving method for activity recognition that enhances the accuracy of activity recognition locally, eliminating the need to stream user data to the cloud. The paper’s contributions are twofold: a Temporal Decoupling Graph Depthwise Separable Convolution Network (TD-GDSCN) to address the challenges of real-time performance and a data augmentation technique to prevent accuracy degradation in real-world environmental conditions. The experimental results show that the TD-GDSCN and data augmentation techniques outperform existing methods in addressing real-time performance and degradation challenges on the NTU-RGB+D 60 and NW-UCLA datasets.