Lightweight Uni-Domain Model for Human Activity Recognition Using UWB Radars

Detecting human activities plays a critical role in many automated systems. It finds a particular application in smart homes with the aim of helping older adults. Building such systems involves overcoming several hurdles such as maintaining privacy, improving user comfort and acceptance, reducing total costs, and ensuring that the utilized models are efficient enough for deployment on edge devices commonly used within Internet of Things (IoT) frameworks in smart homes. In this study, we investigate these issues by assessing the performance of lightweight models for recognizing human activities using a single data modality acquired from three ultra-wideband (UWB) radars. Our experiments achieve 99.87% accuracy under 10-fold cross-validation. This result demonstrates that the presented technique can be applied in real-world settings.