Exploration of Self-Learning Radar-based Applications for Activity Recognition and Health Monitoring

Abstract

Sensor-based monitoring has proven effective in many settings for determining people’s well-being and protecting their safety, even in difficult times like the COVID-19 pandemic. In many applications, radio wave-based systems are more versatile than those based on traditional sensors, thanks to non-contact sensing while preserving privacy. Health monitoring and assisted living are two good examples of how such systems are finding widespread usage in everyday applications. Good performance in such complex monitoring and recognition tasks is often achieved via machine learning. In particular, deep learning can aid with feature extraction, algorithm performance optimization, and forecasting. Yet, to learn how to tackle problems effectively, the generated models usually need access to a substantial amount of data. Furthermore, data preparation may be time-consuming and costly, especially when handled by specialists or when required in real-time systems. Few-shot learning techniques overcome these issues by adapting models to self-learn how to extract meaningful information from limited data. This is feasible by leveraging the learning context and previously acquired knowledge. This doctoral thesis is the result of research on the exploration of fewshot learning techniques for radar-based applications in activity recognition and health monitoring. The investigation was performed by constraining the adaptation of radar-based solutions to limited data, ensuring the robustness of context generalization. The primary goal has been to investigate the use of limited data in very different non-contact applications, each with its own constraints and requirements. Millimeter-wave radar technology and fewshot learning have been used for hand gesture recognition, people counting, and human respiratory signal estimation. Such use cases, ranging from the millimetric displacements of vital signs to the distance of moving individuals, require specific information preprocessing. The generalization learning strategy has been explored for context and user adaptation while also accounting for preprocessing. Some of the algorithms were adapted to run on edge devices, allowing for end-to-end performance estimation and adaptation. The research has been carried out under a doctoral contract at the facilities of Infineon Technologies AG, at its headquarters in Munich, Germany.