Breathe Deeply: Why a Russian Neural Network Needs to Monitor Your Chest

Imagine you're feeling under the weather and your doctor asks you to measure your breathing rate. In the classic scenario, you'd either have to count your breaths yourself (which immediately disrupts your natural rhythm), or get strapped down with sensors, clip on a chest strap, or stick special cannulas up your nose. All of this is uncomfortable, archaic, and causes unnecessary stress. While our smartwatches have learned to count pulse and steps virtuously, breathing has long remained a "contact sport." Researchers from the Russian company "Kryptonite" decided that your laptop's webcam is already smart enough to handle this task without any extra contact.

The idea of non-contact health monitoring isn't new, but it's been, to put it mildly, finicky. Previous attempts to teach AI to "see" breathing often ran into harsh reality: a human is not a mannequin. All it took was a patient scratching their nose, adjusting their glasses, or simply changing position in their chair, and the algorithms would fail. A 13% margin of error in medicine isn't just a statistical inaccuracy—it's the difference between "everything is fine" and "call the ambulance immediately." The problem was that systems couldn't effectively filter natural body movements from the micro-displacements of the chest.

Developer Alexey Protopopov proposed an elegant solution to this problem using the MediaPipe neural network. Instead of simply tracking pixel changes in the frame, the algorithm performs deep image segmentation. It isolates specific zones—the chest and abdomen—and creates a dynamic mask for them. This process is quite resource-intensive and takes about 90% of the total video stream processing time, but it's worth it. Thanks to this "mask," the system knows exactly where to look for the signal, even if you decide to gesture actively or turn sideways to the camera.

The method was tested on a group of 14 volunteers of different genders and ages—from 20 to 65 years old. In total, the neural network "watched" over 2.5 hours of video recordings where people behaved completely naturally. The results confirmed: segmentation allows the system to almost completely ignore interference from posture changes. This is a fundamental difference from old methods that required patients to sit still like in a passport photo. Now the accuracy of RR (respiratory rate) measurement has come very close to the readings of professional medical equipment.

Why is this important right now? Telemedicine is experiencing a boom, but it's still limited by the quality of data patients can send to their doctor from home. If an ordinary office "eye" of a laptop turns into an accurate diagnostic tool, the barrier to quality healthcare falls. We're moving toward a future where your computer will notice signs of shortness of breath or an oncoming illness before you even feel unwell. And it will do this as unobtrusively as possible, without rubber straps across your chest.

Of course, when a neural network starts watching our body movements so closely, privacy questions arise. But from a technical standpoint—this is a major victory of signal processing over the chaos of everyday life. Perhaps soon the phrase "take a deep breath" will sound not from a doctor in their office, but in a notification from your browser.

The bottom line: "Kryptonite" proved that body segmentation with neural networks turns an ordinary camera into a medical device resistant to user movements. Will this become the standard for home check-ups in the next couple of years?