Nvidia Patents a Dashboard Camera That Measures Your Eyelids to Catch Drowsy Driving
Nvidia has filed a patent for a car camera that doesn't just look at your eyes. It measures how far open each eyelid is, in real units, even while your head is tilted or turned.
What Nvidia's dash camera does with your eyelids
Most driver-monitoring cameras judge whether your eyes are closed from a flat 2D image, and a flat image gets confused when you tilt your head up or down. Your eyes can look half-shut when they aren't, or one eye can drop out of view entirely.
Nvidia's patent adds depth to the picture. A camera spots points around your eyes, and the system fits them to a 3D model of a head. That model can even be personalized, built by having you turn your head around once while a depth sensor watches. Depth data then pins those points to real positions in the cabin, so the car can measure how open each eye is in actual units, down to a fraction of a millimeter.
From there the car can count your blinks, notice slow ones, tell when you're looking away from the road, and alert you or take action if you seem to be nodding off.
… generate second 3D coordinates in a global coordinate system for the one or more keypoints based at least on the first 3D coordinates and depth data for the one or more keypoints in the global coordinate system to determine a state of the at least one feature; …
Translation: It combines camera angles and depth measurements to figure out what the driver is doing.
How a flat photo becomes a 3D eyelid reading
The system starts with an ordinary in-cabin camera that finds keypoints (landmarks such as the corners of your eyes, your eyelids and your chin) in each frame. Those are only flat 2D positions, so the next step fits them to a 3D head model using a standard geometry method (Perspective-n-Point, which works out how a known 3D shape must be rotated and shifted to line up with its 2D photo).
Then depth data comes in. It can come from a depth sensor synchronized with the camera, from two sensors compared against each other (stereo), or, when neither exists, from an estimate: the system fits a plane through the eyelid and chin points and traces a ray from the camera to it. The result is a set of 3D positions in a global coordinate system (a shared map of the cabin with one fixed reference point).
From those positions the system can work out:
- Eye openness: the distance between top and bottom eyelid points, checked per eye against that eye's own threshold.
- Blink count, rate and duration, plus PERCLOS (the share of time your eyes are mostly closed).
- Gaze direction, from the pupil's position relative to the eye corners.
To save computing power, the filing also describes switching to the cheaper flat-image eye aspect ratio (eye height divided by eye width) when your face points straight ahead. The same 3D readings can auto-label video for training, replacing people who mark each blink by hand. The claims also cover hands.
The 2D image coordinates from the image data may be converted to 3D coordinates in a global coordinate system using depth data and a 3D occupant model, which may be personalized to the vehicle occupant.
Translation: Regular camera pictures are turned into a 3D computer model that adapts to fit your specific face.
Why head turns stop fooling the eye check
The practical payoff is a drowsiness or distraction alert that works from a real measurement of your eyes instead of a rough guess from a flat image. That matters most in the moments cheaper checks struggle with: when you look down at a screen, tilt your head, or squint into low sun.
The per-eye, per-person thresholds are a useful detail, because people's eyes differ and the filing says the thresholds adjust over time. The auto-labeling idea also matters behind the scenes. Training these systems today means humans watching video frame by frame, which the filing calls expensive and inconsistent. Replacing that with depth-based labels could make the detection models cheaper to build and more consistent.
Nvidia's 82nd filing we've tracked since May in the self-driving sensing race adds to its earlier patents on two-camera distance measuring and depth from regular video.
If you drive tired, the benefit is an alert based on what your eyes are actually doing, not a guess from a flat picture. You would notice it most when you glance down or tilt your head, the moments when simpler checks lose track of your eyes.
The catch is setup. A personalized head model means a short routine of turning your head for the camera, and the filing also describes a cheaper shortcut that skips the 3D math when you face forward. That hybrid looks like the piece that decides whether the idea is practical in a car.
The filing gives no accuracy comparison, so it promises nothing about fewer false alarms. The logic still holds up: a measurement in real units beats a ratio taken from a flat image. For a safety feature, that is a sensible upgrade.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Nvidia patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
14 drawing sheets from US 2026/0311498 A1 · click any drawing to enlarge
Show the other 6 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in