Researchers have developed an AI model that measures your pulse via camera – and for the first time it works reliably when you are moving. No chest strap, no sports watch, no skin contact. Just a face video, an algorithm, a heart rate.
Sounds like science fiction – but it is the state of research as of February 2026. The model is called HBP-Net, published in the journal iScience. What it can do, where it still falls short, and whether it will ever move from the lab into your training – that is what this article is about.
The previous problem with rPPG in training
rPPG (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.
During exercise this breaks down. Head movements, sweat, changing light – all of it creates artefacts that overlay the signal. Classic methods like CHROM or POS reach error rates of 10 bpm and more under motion. That makes them unusable for zone-based training.
What HBP-Net does differently
The conceptual difference: previous rPPG models try to reconstruct the full blood volume pulse signal and then count 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).
What is rPPG?
Remote photoplethysmography uses standard cameras to detect pulse-driven changes from facial colour. Blood absorbs green light more strongly than unperfused tissue – this produces a measurable colour shift with every heartbeat that an AI can evaluate. No electrodes, no strap, no skin contact.
What the study actually shows
The researchers tested HBP-Net on three datasets. For resting measurements (UBFC-rPPG): 0.64 bpm mean error – good, but not the best. RhythmFormer reaches 0.50 bpm there. The difference becomes relevant under motion: on the newly introduced ZJXU-MOTION dataset, which explicitly tests walking and running scenarios, HBP-Net holds up considerably better than PhysNet and remains comparable to or better than RhythmFormer.
Method | MAE at rest (bpm) | MAE walking, low light (bpm) | Parameters |
|---|
HBP-Net | 0.64 | 1.788 | ~0.54 million |
RhythmFormer | 0.50 | comparable | ~3.25 million |
PhysNet | ~1.0 | 3.2 | ~0.77 million |
POS (classic) | 2.44 | > 10 | – |
What this means for training – and what it does not
Motion robustness is the relevant point. If rPPG reaches 1–2 bpm error during running or HIIT, that is in principle usable for training zone management. Current wearables also have 2–5 bpm error on wrist optical tracking at high intensity – the advantage lies more in comfort (no hardware on the body) than in accuracy. The authors see long-term potential for HRV analysis in the heartbeat probability curve too, but that has not yet been implemented.
At the same time: the test setup is still far from real training conditions. 20 participants in the motion dataset, controlled indoor light, no outdoor test. Anyone hoping to upgrade their HIIT monitoring with this right now should know that this is the state of research – not a finished product.
What the study leaves open
Three points remain unresolved. First: how does the model behave across different skin tones? The training dataset is demographically narrow – a known problem in rPPG research that HBP-Net does not solve. Second: outdoor performance. Direct sunlight and changing natural lighting conditions are entirely missing. Third: no code, no dataset download. The study is open access, the model is not.
Source
Yu et al. (2026). HBP-Net: Probabilistic heartbeat detection for robust remote heart rate estimation. iScience, 29(3), 114974. For a broader look at how far AI currently carries in fitness training, see the overview of AI in fitness in more detail.
