-

DSM systems like IRIS should meet very high safety thresholds, typically >95% overall accuracy, extremely low false-negative rates, and tightly controlled false positives. Compliance with ISO 26262 (functional safety) and ISO 21448 (SOTIF) should be mandatory, along with proven reliability across diverse demographics and real-world conditions.

For commercial deployment in safety-critical contexts, fatigue detection systems should meet ≥95% sensitivity and specificity, ≤5% false-negative and false-positive rates, validated across diverse real-world conditions, with continuous post-market performance monitoring.

High level of performance with accuracy and reliability as high as the defense line of my football club

Because IRIS is a safety system used while driving, it must achieve very high accuracy and reliability before commercial release. It should detect drowsiness correctly in most situations while avoiding unnecessary false alerts. Just as important, it must work consistently for different drivers and environments. Strong real-world testing is essential because even small errors could affect safety.

Commercial fatigue-detection systems should meet safety-critical standards with ~95%+ accuracy, false negatives below 1–2%, controlled false positives, and compliance with ISO functional safety and real-world validation requirements.

Better system mobility on all races and skin complexions

The Bottom Line: Success in any project relies on clear communication and consistent action rather than waiting for the "perfect" moment. By breaking down complex goals into manageable pillars—such as prioritizing high-impact tasks, maintaining a feedback loop with your team, and focusing on incremental progress—you create a sustainable workflow that prevents burnout. Ultimately, the goal is to balance efficiency with quality; as long as the core objective remains the "North Star," small daily adjustments will naturally lead to the desired outcome.

Good resulting tests on sample groups of people from all backgrounds, that have a negligible rate of false positives/negatives.

Given the safety-critical nature of Driver State Monitoring (DSM) systems like IRIS, regulatory and industry expectations should be set significantly higher than typical consumer AI applications. Both false positives and false negatives carry risks: false positives may lead to unnecessary driver distraction or system disengagement, while false negatives may directly contribute to road accidents. Although the EU AI Act does not prescribe explicit numerical thresholds for accuracy, it requires high-risk AI systems to achieve a level of performance that is appropriate to their intended purpose and foreseeable risks. For IRIS, this implies: High sensitivity (recall) for detecting genuine drowsiness, to minimise false negatives. Acceptable specificity, to avoid excessive false alerts that may cause alert fatigue. Robust performance across environments, including low light, occlusions (glasses, hats), and varied camera angles. Consistent performance across demographic groups, with minimal disparity between protected characteristics. In practice, commercial deployment should align with automotive safety standards (e.g. ISO 26262, ISO 21448 – Safety of the Intended Functionality) and internal thresholds defined through rigorous validation testing. Regulators are likely to expect documented trade-offs between false positives and false negatives, rather than perfect accuracy.

False negative

False negative

Only systems that demonstrate high accuracy, minimal missed detections, controlled false alarms, and compliance with safety standards should be approved for commercial use.

For a system like IRIS, “high accuracy” is not sufficient. Commercial deployment should only be permitted

More data needs to be collected to correct the issues of false positives.

An excellent level of regulation should he deployed into these areas where Artificial intelligence could make or mar people's lives.

Quality of materials/programs used, Quality Control

For commercial deployment, the system should be required to be as accurate as possible. Deploying a software that has a high probability of making errors would be catastrophic to road safety and thr protection of customers/users of the software.

If the system fails to detect if someone is fatigued then the system could take full control of driving without the users intent. It is my understanding that the vehicle should only take control if a certain situation happens to the user. To help combat this I recommend using multiple metrics to try and eliminate this such as ensuring that the recall metric is high, as then the system will be able to retrieve the data correctly.

Until there is extended proven use, not for use on crowded vehicles etc and where real consent is not obtainable.

I would argue that false negative should generally be minimised more aggressively than false positives, because failing to detect a genuinely fatigued driver could contribute to a serious road traffic accident. However, false positives cannot simply be ignored. If a system generates frequent unnecessary warning, drivers may become annoyed, lose trust in the system, or begin to ignore or disable it.

Real-world deployment should require very high recall for dangerous fatigue events, with acceptable false positives only if they do not create unsafe or confusing interventions. Independent validation across diverse groups is necessary before launch.

Standards should focus on minimizing dangerous misses while keeping false alerts low enough that drivers do not ignore warnings. Testing should be performed on diverse real-world populations, not just internal datasets.

Deployment standards should require a validated minimum on recall, precision, and subgroup parity. Safety-critical systems should be tested against a wide range of realistic conditions.

Deployment standards should be strict enough that missed fatigue events are rare, because the safety impact is severe. False positives should be monitored so the system does not become ignored or disabled.

Performance standards should require evidence of fairness across subgroups and evidence that warning thresholds are safe in practice. High-stakes systems should not rely on untested assumptions.

A real deployment should be validated across demographics, camera conditions, and driving scenarios before use. If subgroup performance is poor, deployment should be delayed.

Deployment should require independent testing, not just vendor claims. The system must prove that it works reliably under different lighting, camera angles, and demographic conditions.

Deployment standards should require a robust evidence base across multiple sites and user groups. A single lab benchmark would not be enough for a safety-critical system.

Deployment standards should require a validated minimum on recall, precision, and subgroup parity. Safety-critical systems should be tested against a wide range of realistic conditions.

A real deployment should be validated across demographics, camera conditions, and driving scenarios before use. If subgroup performance is poor, deployment should be delayed.

Motion eye detection or hands motion detection

The standards should be very high because someone could die if it goes wrong. In an emergency someone may look I'm fit to drive, but their life may depend on it.

For a real deployment of IRIS, I believe that there would need to be a sort of “override option” 4 if the decision made by the recognition software truly is incorrect, i.e., detecting fatigue when none exists. As long as the recognition system is returning false positives and false negatives, its verdict cannot be taken as accurate enough to limit drivers ability to continue driving should they wish.

safer to provide warnings only -> human to verify fact accordingly. Of course, too many false alarms would render the system useless.

From a technical viewpoint, it is important to determine a threshold that can be considered accurate within a strict significance level.