Warning a tired driver is easy. Being right often enough to be trusted is not.
A fatigue monitor earns its place by being right — a system that cries wolf is taped over by lunchtime. Detecting eye closure and distraction from an in-cab camera is only useful if it holds across every face, sunglasses, night driving and vibration, and warns early enough to matter without nagging.
DomainConnected product, driver monitoring
PlatformsConnected Edge
ScopeIn-cab imaging to graded alert
Binding constraintRight often enough to be trusted, not just sensitive
DisclosureRepresentative programme; customer not named
CONTEXT
Where this started.
A driver-monitoring system watches the cab with a camera and infers drowsiness and distraction from eye closure, gaze, head pose and blink dynamics. The value is a timely warning; the risk is a system so prone to false alarms that drivers disable or ignore it.
The difficulty is variation. Faces differ, sunglasses and clear glasses reflect, night driving means infrared imaging, and a cab vibrates and swings between glare and shadow. A model tuned in a lab drifts the moment it meets that range, and every false alarm erodes the trust the system depends on.
So it is engineered for the cab it lives in: infrared imaging that works day and night, models robust across faces and eyewear, and a graded alerting logic that distinguishes a genuine microsleep from a glance at a mirror.
CHALLENGES
4 problems, named.
Stated before any of them had an answer.
01
False alarms kill trust
A monitor that warns at every mirror-check or shadow is switched off or ignored; the whole value rests on a low false-alarm rate.
02
Faces and eyewear vary hugely
Different faces, sunglasses, clear glasses and reflections defeat a model that learned a narrow range; robustness across people is the requirement.
03
Day and night, glare and shadow
A cab swings from bright glare to darkness; the imaging has to work across all of it, which usually means infrared, not visible light alone.
04
Early enough to matter
A warning that fires as the eyes are already closed is too late; the system has to read the onset of fatigue, not just its arrival.
ARCHITECTURE
How it was built.
The model is the easy half. Working across every face, in glare and darkness, and warning early without nagging is what turns a demo into a device a driver leaves switched on.
CONTRIBUTION
What Faststream did.
The scope of the work, rather than a capability list.
Infrared imaging — a capture path that works day and night and sees through sunglasses, not a visible-light camera that fails at dusk.
Robust face and eye tracking — holding across faces, head pose and eyewear rather than a narrow lab range.
Fatigue inference — reading blink dynamics and eye-closure onset to warn early, not only once the eyes are shut.
Distraction detection — gaze off the road distinguished from legitimate mirror and instrument checks.
Graded alerting — warnings that escalate with confidence, so the system is quiet until it is sure.
Edge inference — running in-cab with no cloud dependence, so the warning is immediate and private.
WHAT WAS HARD
The parts that consumed the schedule.
Rarely the subsystem that sounds difficult.
01
Suppressing false alarms
The central engineering is not detecting closed eyes but not warning on the thousands of harmless glances and shadows; that is where the trust is won or lost.
02
Generalising across people
A system that works for one face and not another is useless in a shared fleet; robustness across the population is hard and essential.
03
Imaging a moving cab
Glare, darkness, vibration and eyewear reflections all attack the imaging; infrared and careful optics are needed before any model helps.
04
Reading onset, not arrival
Catching fatigue early enough to matter means modelling its onset dynamics, which is subtler than detecting a fully closed eye.
OUTCOME
What resulted.
Alerts drivers trust
A low enough false-alarm rate that the system stays switched on and is acted upon.
Works across the fleet
Reliable across faces, eyewear and lighting, not just the development set.
Day and night
Consistent detection in glare and darkness through infrared imaging.
Immediate and private
Edge inference in-cab, warning without a cloud round-trip and without shipping video off the vehicle.
Confidentiality
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, confidential deliverables and commercial terms are not disclosed. Where a detail would identify a customer it is omitted rather than approximated. More is available under a non-disclosure agreement, within the limits the customer has agreed.
Why is a fatigue monitor hard if detecting closed eyes is easy?
Because the product is judged on trust, not on detection. Recognising a closed eye in a clear image is straightforward; doing it reliably across every driver, through sunglasses and clear glasses, in glare and darkness, in a vibrating cab, and warning early without firing on every harmless glance is the real problem. A system that raises false alarms gets disabled or ignored, so suppressing them matters as much as catching genuine fatigue.
02
Why infrared rather than a normal camera?
Because a cab swings between bright glare and total darkness, and a visible-light camera fails at night and struggles with sunglasses. Infrared imaging works across day and night and sees the eyes through many sunglasses, giving the model a consistent input regardless of ambient light. Getting that imaging right is a precondition for the detection working at all.
03
Does the video leave the vehicle?
No — inference runs on edge hardware in the cab, so the fatigue and distraction decisions are made locally and immediately, and raw video does not need to be streamed to the cloud. That keeps the warning instant, avoids depending on connectivity, and keeps the driver's image private, with only events logged for review rather than continuous footage.