Advanced Deep Learning Architectures for Human-Radar Interaction in Indoor Environments

Due to the incredible progress in communication technologies and continual reduction in the size of electronic devices, sensors are progressively permeating everyday life. In indoor scenarios, human sensing has led to the creation of smart devices ranging from the usage of human presence or count to vital sensing information to regulate heating, ventilation, lighting, and entertainment systems for a better user experience while saving significant amounts of energy and CO2 emissions in the process. Additionally, sensors have extended in our daily lives to track and monitor our well-being in the form of sleep tracking, vital sensing, or activity monitoring. Deep Learning (DL) has propelled the maturity and robustness of these use cases and reduced the Go-To-Market time of resulting products. The most dominant human sensing sensors are vision sensors which raise privacy concerns and suffer from limited illuminations or occlusions. Contrary to that, radar systems are privacy-preserving, independent of illumination conditions, and allow extraction of information such as range, velocity, and angle of a target. Most works in radar-based human sensing are limited to automotive use cases and use traditional signal processing or conventional DL architectures designed for vision systems. This work explores the use of low-cost millimeter-wave frequency continuous wave radar sensors for human sensing in complex indoor environments, ranging from presence sensing, people counting, and people activity recognition to vital signs monitoring by leveraging advanced deep learning techniques. The limitations of using conventional approaches due to multi-path reflections, clutter, and ghost targets are identified and addressed by using new training schemes, architectures, and loss functions that result in robust and superior performance compared to previous approaches while being widely acceptable due to small computation footprint of the proposed methodologies.