As in-cabin monitoring systems become more sophisticated, validating DMS, OMS and child presence detection across real-world human behaviour, sensor conditions and edge cases is becoming increasingly difficult.
Ahead of InCabin Europe 2026, Javier Salado, Technical Product Manager at Anyverse, explains how synthetic data and simulation can support in-cabin monitoring validation, reduce the domain gap between virtual and real sensor data, and contribute to future Euro NCAP virtual assessment.
1. Which cabin-monitoring scenarios remain hardest to simulate accurately?
The biggest challenge to simulate in-cabin scenarios is human behavior. Humans are unpredictable and can react very differently to different situations. However, still a car cabin is a controlled environment that makes things a bit easier for simulation.
We can define specific situations that can happen inside the car. This means a limited number of behaviors that can affect passenger and vehicle safety. We can focus on simulating those. Efforts like the Euro NCAP assessment protocols are a great starting point. From those, the non-transient states are the hardest: Drowsiness, microsleep, intoxication, unresponsiveness. And these are also the states where real data doesn’t really help you, because you don’t have reliable ground truth either: drowsiness is annotated subjectively and microsleep is labeled after the fact. In simulation, the state is a parameter we set, not something we infer afterwards. On top of the behavior, you then have everything that breaks a DMS in production — sunglasses, masks, headwear, hands on the face — and, beyond the driver, occupant and child presence in rear seats and footwells.
An additional challenge that may not be directly related to the scenarios themselves is to accurately simulate the different sensors and illumination to reduce the domain gap. In-cabin sensing is mostly NIR, and skin and materials at 940nm don’t behave the way RGB intuition suggests.
2.How can synthetic data help address validation bottlenecks?
Synthetic data is in fact the answer to remove the current validation bottlenecks. As I said the cabin environment is somehow controlled, however if you consider all possible combinations of the noise variables to validate different scenarios: driver and passengers demographics, illumination including day and night, additional behaviors, etc. The cross product of all these result in millions of possible combinations. How many of them can you validate with real data? A dozen? A couple dozens if you have the time and resources. The lack of real data to validate is a reality. And for the states that matter most it is not a resources question at all — you cannot put intoxicated or microsleeping drivers on public roads. On top of that, in-cabin data is faces and bodies, so every real collection campaign carries consent and retention constraints.
Using synthetic data for validating can definitely help. Can we generate those millions of combinations? Technically yes, but the point is not brute force, it is designing coverage: sampling the variable space evenly so the gaps real data leaves — low illumination, demographic variability — are actually filled. That is what we built Anyverse INCABIN to do. And you get two things real data can’t give you: exact ground truth for free, and repeatability, so you can rerun the same scenario changing one variable and know that variable explains the difference. For anything that will eventually be an official assessment, that matters as much as the coverage.
3. What metrics should engineers use to evaluate simulation quality?
Simulation quality is a key aspect to provide high levels of confidence in the synthetic data used at different stages of the validation process. Measuring it with specific actionable metrics is not easy.
First you have to look at the fidelity of the simulation technology. Physically correct simulation is a necessary but not sufficient element for high simulation quality. Anything that ensures physically correct simulation, is worth considering.
However, we have to look at the results of the simulation, the synthetic data generated. How can we measure its quality? At the end of the day, synthetic data quality is determined by the perception system that uses it. If the system understands it as if it were real, the synthetic data is of the highest quality. So, it all boils down to measuring the domain gap between real and synthetic data. Lower domain gap, higher quality.
The domain gap has two distinct components:
● Image structure and texture gap: Simulation should reproduce the real sensor output consistently and as accurately as possible.
● Content gap: Simulation should represent real simulations consistently and as accurately as possible.
To measure both aspects we need different metrics. We propose that those metrics are basically kernel distances between different real and synthetic datasets feature spaces:
● Wavelet features for image structure and texture gap
● Foundational computer vision models (DINOv2, CLIP) features for content gap
● System under Test (SuT) features for both at the same time.
What are the acceptable thresholds for the above metrics that define high quality? We find them by correlating the metrics with the systems under test performance: When the SuT performance of the synthetic dataset is close enough to the performance of the real dataset, the domain gap metrics values can be considered acceptable.
4. How close are virtual validation methods to regulatory acceptance?
We are getting closer, and it is worth being precise about where we are. DMS is already regulated and Euro NCAP already assesses it today, physically. What is ahead is virtual assessment, which Euro NCAP is clearly pushing for their future assessment protocol in 2029. And simulation is not new to regulators — UN R157 already accepts it as part of the audit path for automated lane keeping. So the question is not whether virtual validation can be accepted in principle. It is what evidence a DMS or OMS has to provide, and we still have to build trust in the synthetic data and define the protocols to use it in the official assessments.
At Anyverse we have put together the Virtual Assessment Implementation Program (VAIP), bringing together different industry stakeholders to build that trust and propose implementable protocols that Euro NCAP’s framework can adopt — the framework is theirs, how the industry implements it credibly is still open. The first goal of the program is precisely to define the above metrics and the thresholds that build confidence on the synthetic data. It is open to OEMs, Tier-1s and test houses working on DMS, OMS and CPD, and anyone at InCabin who wants to join can reach us at https://anyverse.ai/virtual-assessment-implementation-program/
Interested in exterior vehicle sensing technology?
With a pass to InCabin Europe, you’ll also get full access to our
co-located sister event, AutoSens. Find out more here >>
