An interview with Gabriel E. Hine, Smart Eye, ahead of his session at InCabin Europe.
Iris recognition has long been one of the most accurate ways to confirm who a person is. Yet for decades, it stayed locked inside smartphones, border-control gates, and dedicated access terminals.
Ahead of his session at InCabin Europe, Gabriel Hine, an Iris Authentication Technology Owner at Smart Eye explains how his team moved iris recognition out of those controlled settings and into the vehicle cabin, why combining iris with face recognition beats either one alone, and what this unlocks for the car once identity can be established from cameras the vehicle already carries.
1. What technical breakthrough enabled reliable low-resolution iris recognition?
The short answer is that we stopped depending on a single, perfect image.
Traditional iris recognition was built around very demanding optics. The relevant ISO standard effectively calls for around 17 pixels per millimeter on the iris, which works out to roughly 200 pixels across the iris diameter. You get that from a phone held at arm’s length or a kiosk you lean into, but you do not get it from a driver-monitoring camera sitting on the steering column 60 to 120 centimetre’s away. In the cabin, the same iris might be covered by only around 36 pixels, close to an order of magnitude less detail, on top of variable lighting, head movement, eyewear and motion blur.
Two things made low-resolution iris recognition work in that environment.
The first is temporal integration. Instead of trying to extract a full iris signature from one frame, which breaks down well before you get down to 36 pixels, we aggregate information across many frames of the video stream. A cabin camera is running continuously, so what looks like a limitation (each individual frame is poor) becomes an advantage (there are a lot of frames). By combining that evidence over time, we build an iris representation that is far more discriminative than anything a single low resolution frame could give us.
Crucially, “many frames” does not mean a slow system. Cabin cameras typically run at 30 or 60 Hz, so a large number of frames are captured within a very short window. In practice that means enrolment takes around two seconds, and an authentication check completes in under a second, fast enough to feel instant to the driver even though it draws on a whole sequence of images rather than a single snapshot.
The second is less glamorous but just as important: Smart Eye has spent decades doing eye tracking on imperfect automotive imagery. Before you can recognize an iris, you have to find it, and finding the eye, the iris and the pupil reliably under real driving conditions is genuinely hard. That localization and segmentation know-how, built up over years of developing driver monitoring, provides a foundation that conventional close-range iris systems never needed, since they could assume a cooperative user and a clean image.
Together, temporal integration and this automotive eye tracking expertise let us run iris
recognition at the resolution that standard occupant monitoring cameras already deliver, with no new sensor in the car.
2. How does multimodal fusion improve authentication accuracy?
Before getting into the “why”, it helps to define two terms, because they drive everything in biometrics, and most people outside the field have never had to think about them.
False Acceptance Rate (FAR) is how often a system wrongly says “yes” to the wrong person: an impostor gets in. False Rejection Rate (FRR) is how often it wrongly says “no” to the right person: you are yourself, but you get turned away. These two always pull against each other. Tighten the system to make impostors almost impossible (a very low FAR), and you start rejecting legitimate users more often (a higher FRR). The art is getting both low at the same time, and that is exactly where combining modalities helps.
Face and iris have complementary strengths, so fusing them lets us take the best of each:
- Iris keeps working when the face is obscured. If someone is wearing a face mask or a large part of the face is covered, iris still has a clear signal to work with.
- Face keeps working when the iris signal is weak. At lower resolution, at an awkward gaze angle, or with certain types of glasses, the face is often the more reliable of the two.
Because the two modalities tend to fail in different situations, a combined system stays confident across far more real-world conditions than either alone, and it can lean on whichever signal is stronger at that moment. The net effect is a lower false-rejection rate at any given security level. For the driver, that translates directly into a more user-friendly experience: they can simply behave naturally and are not asked to remove a face mask, take off glasses, or hold still and stare at the camera to be recognized. If one trait is momentarily weak, the other can complete the check.
There is a second benefit that matters just as much: not every task in the car needs the same level of security, and fusion lets you adjust the trade-off to the task.
- Convenience features can run at a more relaxed security level. Recognizing that the driver has changed, or loading someone’s seat, mirror and infotainment
preferences, is low-stakes. If the system occasionally gets it wrong, nobody is
harmed, so a face-only check or a modest iris signal is perfectly adequate. Here you
can accept a higher FAR in exchange for a fast, seamless, effortless experience. - Payments and high-value actions demand a very low FAR. Authorizing an in-car payment or unlocking sensitive data is exactly where mistakenly accepting an impostor is costly, so you want the strongest possible assurance. In that situation the system can ask the driver to cooperate briefly with the camera so it captures a good sample of both traits, and then require both to agree.
The reason demanding both traits is so powerful comes down to statistics. Iris and face are essentially independent biometrics: iris patterns are not written into our DNA and carry an enormous amount of natural randomness. This is why iris recognition can even tell identical twins apart, despite their faces being nearly impossible to distinguish.
Because the two biometrics are independent, the odds of an impostor fooling both at once are close to the product of the odds of fooling each one alone. In plain terms, if one trait might wrongly accept an impostor once in a ten thousand attempts and the other once in a hundred thousand, fooling both at the same time is closer to one in a billion. Combining them therefore drives the false-acceptance rate down dramatically.
In essence, the same hardware and the same enrolment serve both ends of the spectrum:
our fusion approach adapts to the level of security the moment calls for, staying effortless for everyday convenience and tightening up to banking-grade assurance when it matters.
3. What challenges arise when interpreting natural driver behaviour?
Biometrics in a car involve something people are rightly protective of: their identity and their body. Getting the privacy model right is not an afterthought, it is a precondition for deployment.
The good news is that we do not have to invent that model from scratch. Android’s tiered biometric model and the FIDO Alliance’s passwordless specifications already encode it. Both were written and audited by a broad group of industry, payment and government stakeholders, and customers rely on them on their phones every day. Between them, these two frameworks imply a handful of principles that carry directly into the cabin:
- Process on the device, and store templates rather than images. The sensitive work happens locally, inside a protected part of the system. What gets stored is a mathematical template, not a photograph of your eye or face, and the secure pipeline is designed so that raw data cannot be quietly read out or synthetic data injected.
- Keep authorization local. The biometric check happens on the device to release a cryptographic key, and the biometric itself never leaves. What is exposed to the outside world is a cryptographic assertion, not your face or iris, exactly the way a passkey works.
- Match the assurance to recognized security tiers. The tiers exist precisely so that a payment or a keystore-backed action is only ever authorized by a system that has proven both its accuracy and its resistance to spoofing. This is also what lets in-car biometrics align with regulations that banks already live under, such as Europe’s Strong Customer Authentication rules.
- Defend against spoofing, transparently. A biometric system has to assume people
will try to fool it, with a printed photo, a replayed video, a mask, or a printed contact
lens. Presentation-attack (spoof) detection is a core part of the pipeline, and
different levels of robustness are defined and independently tested against recognized standards. Just as important is being clear with occupants about what is captured, why, and how it is protected.
The overarching point is that none of this requires the driver to trust a black box: the protections map onto the same standards customers already rely on every day.
4. What applications beyond authentication become possible?
Authentication is the obvious starting point, but identity becomes much more useful once it is available continuously and without asking the driver to do anything.
The first application is personalization. If the car knows who has entered, it can load the right seat and mirror positions, climate settings, navigation history, infotainment preferences, and accessibility features before the driver has touched a control. In a shared vehicle, that turns the cabin from a generic space into one that adapts immediately to the person using it.
Identity can also help control access to data and services. A vehicle may contain payment credentials, personal messages, work accounts, or AI services that should not be available to every occupant. Instead of treating the entire car as either locked or unlocked, the system can decide which features and information each person is allowed to access.
There are safety applications as well. Knowing who is driving can help the vehicle apply the right permissions, restrictions, or support. A young driver, a fleet employee, and the vehicle owner may all be allowed to use the same car, but not necessarily in the same way. The identity layer gives the vehicle a reliable basis for making those distinctions.
The longer-term opportunity is continuous identity. Traditional authentication usually happens once, at the start of a session. In a vehicle, the camera is already active, so the system can continue checking that the person who was authenticated is still the person in control. That could help protect payments, sensitive actions, or access to personal data throughout the journey rather than only at the moment the car is unlocked.
In that sense, iris and face recognition are not just another way to replace a key or a password. They give the vehicle a persistent understanding of who is present, which opens the door to a much broader set of personalized, secure, and context-aware services.
Interested in exterior vehicle sensing technology?
With a pass to InCabin Europe, you’ll also get full access to our
co-located sister event, AutoSens. Find out more here >>
