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

Area of Science:

• Engineering
• Computer Science
• Human-Computer Interaction

Background:

• Privacy concerns limit traditional vision-based and wearable sensor methods for human activity recognition in sensitive environments.
• Non-invasive and anonymous monitoring is crucial for applications in settings like bathrooms.

Purpose of the Study:

• To investigate the feasibility of using Doppler radar for human activity recognition in privacy-sensitive bathroom environments.
• To develop and evaluate deep learning models for classifying daily living activities using radar data.

Main Methods:

• Utilized a BGT60TR13C Xensiv 60 GHz radar sensor for data collection.
• Collected a dataset of ten daily living activities from seven volunteers in a bathroom setting.
• Employed deep learning models, including DenseNet201 and ResNet50, with bidirectional long short-term memory networks for activity classification.

Main Results:

• Achieved high overall accuracy, with DenseNet201 reaching 97.02% and ResNet50 reaching 94.57%.
• Demonstrated strong recognition performance for most activities, including face washing, teeth brushing, and dressing/undressing.
• Identified 'lying down' and 'getting up' as challenging activities due to motion similarity.

Conclusions:

• Doppler radar-based human activity recognition is a viable and privacy-preserving alternative to invasive monitoring systems.
• The proposed radar approach offers an effective solution for smart home and healthcare applications requiring non-invasive monitoring.