Charlie replied: For sure, also, for an EU vehicle deployment, the main hard requirements would come from the General Safety Regulation framework and the specific DDAW rules. EU Regulation 2019/2144 requires Driver Drowsiness and Attention Warning systems for M and N category vehicles, with the obligation applying to new vehicle types from 6 July 2022 and to all new vehicles from 7 July 2024. DDAW systems are defined as systems that assess the driver’s alertness through vehicle-system analysis and warn the driver if needed. The detailed DDAW regulation does not prescribe one model architecture or one fixed accuracy percentage. Instead, it sets functional and validation requirements. The system must operate under defined conditions, including automatic activation above 70 km/h and operation in daytime and night-time conditions. It must warn the driver at a drowsiness level equivalent to or above level 8 on the Karolinska Sleepiness Scale, although it may warn from level 7 onwards. The validation regime is also important. Manufacturers must compile a dossier explaining the DDAW system, its operation, the test procedures used, the rationale for those procedures, and the full validation results from human-participant testing. The technical service then assesses whether the design, operation, and performance evidence adequately demonstrate compliance, and it may run a manufacturer-defined verification test. So, legally, the answer is: IRIS must meet the DDAW type-approval performance requirements, but those requirements are not expressed as a simple public accuracy percentage. The regulation requires validation against drowsiness states, using the Karolinska Sleepiness Scale or an equivalent method, with human participants and a statistical approach to minimum performance thresholds. At least 10 human participants must be included in validation testing, and each participant must generate at least one true positive or false negative event.

Deleuze replied: Definitely the case. It is also correct, as some other commenters have mentioned, that standards such as ISO 26262 and ISO 21448/SOTIF are also highly relevant. ISO 26262 applies to safety-related electrical and electronic systems in production vehicles, while ISO 21448 is especially relevant where safety depends on complex sensors and processing algorithms. These standards require a safety case showing that risks have been identified, reduced, verified, and validated, but they also do not impose a single drowsiness-detection accuracy percentage. Euro NCAP is another important commercial benchmark. It is not a legal approval regime, but it strongly influences market expectations. From 2026, Euro NCAP evaluates driver monitoring technologies that maintain attention and engagement, with points awarded for systems that monitor driver behaviour in real time and link driver-state information to assistance-system behaviour.

Charlie replied: Good for projects but not useful advice for specific standards for an AI algo imo

Deleuze replied: It should also be documented! The system should document, at minimum: drowsiness detection recall/sensitivity at KSS level 8 and above; false negative rate, because missed fatigue is safety-critical; false positive rate or false alerts per driving hour, because excessive false alarms reduce trust and may cause users to disable or ignore the system; time-to-warning once drowsiness is detected; performance across demographic groups, lighting conditions, occlusions, glasses, hats, skin tones, facial hair, and camera angles; worst-performing subgroup results, not only average results; confidence intervals and statistical significance; system availability, failure detection, and fallback behaviour; post-market monitoring results and drift over time.