Radar-Based Activity Recognition in Strictly Privacy-Sensitive Settings Through Deep Feature Learning

Abstract

Human activity recognition in privacy-sensitive environments, such as bathrooms, presents significant challenges due to the need for non-invasive and anonymous monitoring. Traditional vision-based methods raise privacy concerns, while wearable sensors require user compliance. This study explores a radar-based approach for recognizing the activities of daily living in a bathroom setting, utilizing a BGT60TR13C Xensiv 60 GHz radar, manufactured by Infineon Technologies AG (Munich, Germany, EU), to classify human movements without capturing identifiable biometric features. A dataset was collected from seven volunteers performing ten activities which are part of daily living, including activities unique to bathroom environments, such as face washing, teeth brushing, dressing/undressing, and resting on the toilet seat. Deep learning models based on pre-trained feature extractors combined with bidirectional long short-term memory networks were employed for classification. Among the 16 pre-trained networks evaluated, DenseNet201 achieved the highest overall accuracy (97.02%), followed by ResNet50 (94.57%), with the classification accuracy varying by activity. The results highlight the feasibility of Doppler radar-based human activity recognition in privacy-sensitive settings, demonstrating strong recognition performance for most activities while identifying lying down and getting up as more challenging cases due to their motion similarity. The findings suggest that radar-based human activity recognition is a viable alternative to other more invasive monitoring systems (e.g., camera-based), offering an effective, privacy-preserving solution for smart home and healthcare applications.