Driver Fatigue Cameras: Promising Trial Results, Thin Independent Proof, Rising Privacy Claims

The use case

A carmaker's built-in driver-attention system detected about 10% of severe impairment events in an Australian test, with a 33% false-alarm rate, while a dedicated fleet camera system performed far better (measured in a controlled test, as summarised by a motoring club in March 2025)¹. That gap frames the whole use case: cameras that watch the driver are not all equal, and few of the differences have been tested in public.

In a typical fleet product, a camera faces the driver, software reads eyelid closure, head position and gaze, and an alarm sounds in the cab when the pattern suggests drowsiness. Many products also send the event to a manager³⁵.

What counts as AI. Here AI means machine-learning models that interpret camera images. Samsara states that its drowsiness model was trained on more than 38 billion minutes of video (vendor-reported, October 2024)⁵. None of the sources opened confirm machine learning for the other products discussed, so that point is left open. Older technologies do similar work and are not credited as AI: camera-based eye-closure measurement was already evaluated in a US Department of Transport report covering 2005 to 2006¹², and EU law defines the basic drowsiness warning as assessing alertness through vehicle systems analysis, which does not require a camera⁷. Wearables, steering-wheel sensors and fatigue-management rules are likewise out of scope. Phone, seat-belt and collision detection are mentioned only where they bear on the evidence.

Where it is running

The car figures cover all driver and occupant monitoring, not only fatigue, and the company links the rise to European rules⁶. Outside the United States and Australia, no named fleet deployment with published results was found.

What the evidence shows

The strongest independent evidence is an Australian research programme by Monash University, Central Queensland University and the Institute for Breathing and Sleep, funded with A$1,106,410 by the Australian Automobile Association (final report, May 2024)². Only the opening of the report was read directly; the results below come from a summary by the NRMA motoring club (March 2025), which also carries promotional text for its own driver training¹.

• Randomised trial with fleet drivers. Alarms from the Optalert system were linked to more than 50% fewer critical drowsiness events, and alarms from Seeing Machines Guardian to 55% fewer fatigue events, 48% fewer hard-braking events and 34% fewer hard-cornering events (randomised trial, as summarised by NRMA)¹.
• Controlled testing of alarm quality. A single Guardian alarm was linked to 6 times (mild), 19 times (moderate) and 27 times (severe) impairment, with severe meaning a near-crash within five minutes; the summary does not explain the baseline (controlled test, as summarised by NRMA)¹.
• Weaker systems. The built-in Subaru system detected about 10% of severe events and none of the mild or moderate ones; a steering-wheel heart-rate device produced no fatigue alarms (controlled test, as summarised by NRMA)¹.
• Older trial. A vendor page describes a 2017 peer-reviewed study in which a driver fatigue alert reduced fatigue events by 66.2%, and adding feedback to the employer raised the reduction to 94.4% (vendor-reported description; the paper could not be opened and is older than three years)³.

The mechanism matters. In the 2017 result, the larger gain came from a person acting on the event, not from better detection³. A reduction in flagged events also depends on how the system defines a validated event, which the sources read do not show.

Independent comparisons of fleet cameras are scarce. A test-track study by the Virginia Tech Transportation Institute, funded by the vendor Motive, found Motive's camera alerted drivers 86% of the time, Lytx 32% and Samsara 21% across six risky behaviours (test track, March to May 2023; vendor-funded). Drowsiness was not among the six, and the study was not designed to measure false alerts. Lytx said the study did not compare precision⁴. A research prototype reported in January 2026 detected eye closure in 88.89% and yawning in 85.19% of cases with 27 participants (abstract, prototype, not a fleet product)¹³.

What it takes: cost, integration, governance

Cost. No official price per vehicle was found, so none is given. Revenue per unit cannot be derived from the operator figures read⁶.

Integration. The benefit in the 2017 trial depended on an employer response³, and Samsara's product sends alerts to managers by text or email and aggregates reports⁵. A fleet therefore needs a process and staff time as well as hardware.

Rules in the European Union. Regulation (EU) 2019/2144 lists a driver drowsiness and attention warning and an advanced driver distraction warning among required vehicle systems, and applies from 6 July 2022⁷. The systems must not continuously record or retain data beyond what the closed loop needs, must not make data available to third parties, and must delete it immediately after processing. The recitals say these systems should work without biometric information, including facial recognition⁷. The Annex II dates for each vehicle category could not be read directly. Seeing Machines states that the camera-based requirement came into force on 7 July 2026 (operator-reported)⁶. The regulation text read does not mention cameras, and it is not established that it covers add-on fleet cameras that send alerts to managers.

Driver acceptance. A University of Nottingham study for the RAC Foundation surveyed 477 UK drivers in scenarios of monitoring, and reported on 8 October 2026 that 12% found fatigue monitoring completely unacceptable and 18.4% completely acceptable (self-reported survey; press release, report not opened)¹¹.

Where it falls short

The challenge question is what would make this use case look less successful than it is presented. Four points emerge.

Surrogate outcomes. The trial results count alarms and driving events, not crashes¹. No crash-rate result was found.

Human response. Gains that depend on an alarm plus a manager cannot be credited to the camera alone³.

Unverified vendor claims. Vendor pages claim reductions above 90% and large numbers of prevented crashes, but the 94.4% figure came from a study condition that included employer feedback, and the vendor's extrapolation to its own product was not confirmed³. Samsara's figure that 77% of drowsy events were detected by behaviours other than yawning alone is vendor-reported with no accuracy test⁵.

Legal challenges. Illinois privacy law is the most active front in the United States. A federal court in Illinois refused to dismiss a drivers' claim against HMD Trucking over driver-facing cameras in June 2025 (year inferred from the article date), at the pleading stage only⁸. Samsara agreed a US$3.95m settlement in the Karling case, with a final approval hearing set for 16 October 2025; whether approval was granted was not confirmed⁹. In April 2023 the US National Labor Relations Board found that inward-facing cameras in a truck gave the impression of union surveillance (older than three years)¹⁰. These cases concern consent and surveillance, not accuracy.

The evidence also covers few countries. The trial evidence is Australian and the legal record mostly American.

Phase: In early use — why

The hypothesis was In wide use. The evidence supports a lower phase. Driver monitoring in cars is spreading quickly under European rules, but the figures cover all monitoring, not confirmed machine-learning fatigue detection⁶. Fleet camera sales are a few thousand units per year for one supplier, and independent results are limited to one trial programme¹⁶. Because the share of genuine machine-learning use cannot be established, the lower plausible phase is used.

What to watch next

• Audited FY2026 results from Seeing Machines, which the company said would follow before the end of September 2026 (operator-reported); check whether an installed-base figure for Guardian is disclosed⁶.
• A docket or settlement-site confirmation of final approval in Karling v. Samsara after the 16 October 2025 hearing⁹.
• The next ruling in the HMD Trucking case, which survived a motion to dismiss on the Illinois biometric privacy claim⁸.
• Publication of the Annex II application dates for the EU drowsiness and distraction warnings, to confirm which vehicle categories fall under the rule⁷.
• Any independent replication of the Australian trial in another country or fleet, which would test whether the 55% result holds¹.

Questions to ask before adopting

• Has an independent party tested this vendor's system for drowsiness, with false-alarm rates, rather than relying on the vendor's own counts?
• Who receives alerts and video, how long are they kept, and does that comply with biometric and labour law where drivers work?
• What happens after an alarm, such as a rest stop or a manager call, and who is paid to do it?
• Will success be measured by fatigue events, near-misses or crashes, over what period and against what baseline?
• How do drivers view the system, and what is the policy when a camera is covered or disabled?