Christopher Klenk4 min readResearchers have built an AI model that measures your pulse via camera and stays much more stable while walking than earlier models. You do not need a chest strap or a watch, just a camera pointed at your face.
The HBP-Net study was published in the journal iScience in February 2026. I looked at what the model can do, where it still falls short and whether it will ever move from the lab into your training.
At a glance
HBP-Net detects the pulse from face videos by computing, for every frame, the probability of a heartbeat instead of rebuilding the whole pulse curve. At rest it is roughly as good as the best models, during slow walking it is clearly better. It was only tested in the lab and on few people, though. Code and motion data are public, a finished product does not exist.
Stay on top of AI training & analysis? One click — and Google shows you my articles first.
Add as a preferred source on GooglerPPG (remote photoplethysmography) works like this: cameras capture minimal colour changes in the facial skin caused by blood flow. The pulse can be derived from this signal. Under controlled conditions this works surprisingly well. For resting measurements, current models are under 1 bpm error.
As soon as you move, it gets much less accurate. Head movements, sweat and changing light disturb the signal. Classic methods like CHROM or POS already show errors of just over 2 to almost 8 bpm on resting recordings in this study's comparisons. Under motion that tends to get worse, and for training by heart rate zones it is too inaccurate.
The trick: previous rPPG models try to rebuild the full pulse curve and then count the peaks. HBP-Net skips this step. For each frame in the video, the network directly computes a probability that a heartbeat is occurring. Hence the name Heartbeat Probability Net.
The result is a smoother curve that is more robust against noise. Instead of individual sharp spikes, you get broad probability bumps from which the pulse is calculated. With around 540,000 parameters, the model is significantly smaller than competitors such as DeepPhys (~7.5 million) or RhythmFormer (~3.25 million).
Remote photoplethysmography uses standard cameras to detect pulse-driven changes in facial colour. Blood absorbs green light more strongly than tissue with less blood, so your skin colour changes very slightly with every heartbeat, and that is what the AI picks up. No electrodes, no strap, no skin contact.
The researchers tested HBP-Net on three datasets. At rest (UBFC-rPPG) the mean error is 0.64 bpm. That is good, but RhythmFormer manages 0.50 bpm there. It gets interesting on the new ZJXU-MOTION dataset with standing, slow walking and fast walking. Results are only published for standing and slow walking, though. While walking, HBP-Net is clearly ahead of PhysNet and of RhythmFormer too.
Method | MAE at rest (bpm) | MAE walking, low light (bpm) | Parameters |
|---|---|---|---|
HBP-Net | 0.64 | 1.40 | ~0.54 million |
RhythmFormer | 0.50 | 7.79 | ~3.25 million |
PhysNet | 2.95 | 2.61 | ~0.77 million |
POS (classic) | 2.44 | not tested | not stated |
During slow walking, the error of HBP-Net ranges from 1.4 to 6.6 bpm depending on the light. My reading: for heart rate zones on the treadmill, something like this could be enough one day. Whether it holds up during running or intervals was not tested, so for now that is just a guess. I see the advantage more in comfort than in accuracy, since you wear nothing on your body. The authors see the probability curve as a basis for future HRV analysis, but they write themselves that the precision is not there yet.
The test setup is still far from real training. The motion dataset has 20 people, and after a 3 to 1 split about 5 of them were used for testing. Everything happened in the lab, daylight was only used while standing, and there was no outdoor test. If you want to monitor your training with this right now, know that this is research, not a finished product.
A few points remain open. First, skin tone: all 20 people in the motion dataset reported East Asian ethnicity. How the model performs on other skin tones, the study does not show. Second, light: with daylight the error was highest, even while standing (4.3 bpm), and there is no outdoor test. On the plus side, the code on GitHub and the motion dataset are public.
Ye et al. (2026). HBP-net for robust remote heart rate estimation using heartbeat probability. iScience, 29(3), 114974. For more on AI in training, see the overview of AI in fitness.
