Non-Contact Heart Rate Estimation for Moving Subjects via aDual-Trajectory Hybrid Deep Framework

Imagine trying to listen to a single heartbeat in a room where the walls are shaking, the floor is vibrating, and a crowd is moving around. For decades, medical sensors have struggled with this exact problem when they rely on contact. If a person moves, the device slips, the signal breaks, or the motion of the body drowns out the tiny, rhythmic thump of the heart. This limitation has kept continuous, comfortable heart monitoring out of reach for many daily situations, from walking around a hospital to exercising at home. Scientists have long sought a way to measure these vital signs without touching the skin, using invisible waves that can pass through the air to detect the subtle rise and fall of a chest. The challenge has always been separating the faint signal of a beating heart from the loud, chaotic noise of a moving body.

A team of researchers at Northwest University in Xi'an has developed a new approach to solve this problem using millimeter-wave radar, a technology that sends out harmless radio waves to sense movement. Their method works in two distinct stages, acting first like a skilled tracker and then like a sophisticated sound engineer. In the first stage, the system does not simply pick a fixed spot to listen, which would fail if the person moves away from that spot. Instead, it watches the person's movement across a map of space and time, fusing two different ways of tracking their path to find the most reliable route. Once it locks onto this path, it extracts a raw stream of data that captures the chest's movement, even while the person is walking or shifting position. This step is crucial because it ensures the system is listening to the right person, not the background noise or a ghost signal caused by motion.

The second stage takes this raw, noisy stream and cleans it up using a specialized computer model designed to understand the shape of a heartbeat. This model breaks the signal into different layers of frequency, much like separating the bass, drums, and vocals in a song, to isolate the specific rhythm of the heart from the jarring sounds of walking or stepping. It then uses a deep learning network to reconstruct a clear, smooth waveform of the heartbeat, effectively removing the interference that would otherwise make the reading impossible. Finally, the system compares this clean waveform against a known template to count the beats per minute. The researchers tested this system on eight people in a closed corridor, having them stand still, walk forward, and walk backward, while wearing a smart bracelet to verify the accuracy.

The results show that this dual-stage approach works remarkably well in dynamic situations where older methods fail. When the subjects were moving, traditional signal-processing techniques produced errors that were often large and inconsistent, with some estimates missing the true heart rate by more than twenty beats per minute. In contrast, the new method kept its accuracy, with an average error of just over two beats per minute. In tests where the system was applied to different environments, such as a meeting room and a busy office, it maintained a high level of precision, with errors staying below four beats per minute even when multiple people were moving around. The researchers found that their method outperformed several other advanced computer models and standard filtering techniques, proving that combining a smart tracking system with a powerful reconstruction network is a viable path forward.

While the study demonstrates significant progress, the authors note that their current success relies on data they created by simulating motion interference on clean signals, and they have not yet tested the system on a massive, diverse group of people in the real world. They also point out that the system currently works best with a single radar unit and a single person, and that complex scenarios with many moving individuals still present challenges. However, the core finding is clear: by first locating the correct path of movement and then using a hybrid deep learning framework to reconstruct the heartbeat, it is possible to monitor heart rates accurately even when a person is in motion. This breakthrough suggests a future where vital sign monitoring can be truly contact-free and reliable, allowing for continuous health tracking without the need for straps, sensors, or physical attachment to the body.