Multirange Breathing Rate Estimation With Deep Learning Using FMCW Radar
Breathing rate reflects both health status and daily activities of humans. Irregularities in breathing can indicate stress, sleep disturbances, or underlying respiratory conditions. Conventional systems that rely on chest straps or wearable sensors are highly accurate but require direct contact with the body, which could be inconvenient and unsuitable for many. This has driven interest in contactless technologies that can continuously monitor breathing, without any discomfort.
Frequency-modulated continuous wave (FMCW) radar has emerged as a promising technology for contactless monitoring. Unlike optical or camera-based systems, radar can capture minimal movements with high precision and works reliably under various conditions. It operates by sending out a signal that reflects off the chest wall; the returning signal carries subtle shifts corresponding to inhalation and exhalation. However, it is not easy to extract breathing information from these signals. Reflections from nearby objects, background movements, and the dispersion of respiratory signals across several distance ranges make accurate estimation challenging.
Choosing a single range approach, or “bin”, from where the radar signal appeared strongest was effective under ideal conditions. However, it loses information when breathing movements are distributed across multiple ranges, resulting in reduced accuracy, especially in dynamic environments where people or objects may also be moving in the background.
The study introduces a novel solution: a deep learning method that simultaneously combines information from multiple ranges. Instead of narrowing the focus to a single bin, the system integrates signals from several neighboring ranges. This multirange perspective provides a more complete view of the breathing pattern and enhances robustness against noise and interference. Additionally, the model is lightweight and requires less than 25 kilobytes of memory for 30-second input data, making it compact enough for embedded devices designed for everyday use.
To evaluate its effectiveness, the experiments were conducted on 25 participants under a wide range of conditions, including distances between 0.3 and 1.5 meters, angles of up to ±30 degrees, various mounting heights of the radar, and both static and dynamic environments. Reference breathing rates were recorded using a medical-grade device for comparison.
Results demonstrate that the multirange deep learning model consistently outperformed single-range methods. Over 30-second observation windows, it achieved a root-mean-square error of only 1.88 breaths per minute and a strong correlation of 0.83 with the reference. Even with shorter 10-second windows, more suitable for quick response applications, the system maintained accuracy, with an error of 2.39 breaths per minute. The model remained reliable across variations in distance, angle, and background activity, highlighting its adaptability to practical scenarios.
The combination of radar sensing with efficient deep learning has made contactless breathing monitoring accurate and resource-friendly. By capturing a richer set of signals and processing them with a compact model, the system is effective and reliable for everyday monitoring. Potential applications extend from smart homes and workplace wellness to in-vehicle monitoring and sleep assessment. It can further be developed to track vital signs, such as heart rate, or integrated with motion-rejection strategies to handle even more complex scenarios.