Action Recognition from 4D Point Clouds for Privacy-Sensitive Scenarios in Assistive Contexts

Abstract

Dementia is one of the leading causes of disability and dependency among older people worldwide. To address the challenges faced by people with dementia, vision-based technologies have been proposed to provide context-aware assistance. These technologies typically rely on cameras to understand actions and tailor assistance accordingly. However, privacy concerns hinder their adoption, particularly in privacy-sensitive contexts. This study proposes the use of 4D point clouds as a privacy-preserving modality for assistive systems. By relying only on 3D data and excluding RGB information, we aim to enable personalised assistance while mitigating privacy risks. To assess the feasibility of this approach, we collect a real-world dataset with the help of 16 people with dementia and evaluate the state-of-the-art P4Transformer model on this dataset. Our results show promising performance, demonstrating the viability of point clouds as a practical alternative for privacy-sensitive action recognition in real-world settings. However, the model does not achieve the performance attained on benchmark datasets, highlighting the importance of adapting models to deal with the complexity of real-world data. By addressing privacy challenges and validating the model with real-world datasets, this research contributes to the advancement of privacy-aware assistive systems for people with dementia, towards more personalised and effective dementia care.