Multirange Breathing Rate Estimation With Deep Learning Using FMCW Radar

Frequency-Modulated Continuous Wave (FMCW) radar is a promising technology for non-invasive monitoring of vital signs, such as breathing rate (BR). A key challenge lies in selecting the correct range bin corresponding to the human target, particularly in dynamic environments and resource-constrained systems. This paper introduces a resource-efficient deep learning (DL) method for multi-range BR estimation that significantly enhances accuracy while maintaining low computational and memory requirements, making it suitable for embedded systems. We benchmark our method against two alternative single-bin approaches: energy-based and Spatial Phase Coherency (SPC) range-bin selection, both integrated with deep learning for BR estimation. Performance is evaluated on an experimental dataset comprising recordings from 25 subjects, with 17 used for training and 8 for testing. Recordings were conducted under diverse conditions, spanning various distances, angles, sensor mounting heights, and static and dynamic background scenarios, including individuals walking in the background. Our method achieves a Root Mean Square Error (RMSE) of 1.88 breaths per minute (bpm) and a Spearman correlation coefficient of 0.83, outperforming state-of-the-art methods that typically report higher RMSEs on smaller datasets. With a minimal memory footprint of approximately 25 KB, the proposed approach is well-suited for resource-constrained embedded systems. This advancement underscores the potential of lightweight deep learning for reliable and accurate non-invasive vital sign monitoring in healthcare applications.