As AV systems increasingly rely on cameras and sensors to interpret the road environment, long-standing assumptions embedded in road standards are being brought into question. Today’s road markings, signage, materials, and lighting were designed around the limitations of human visual perception, assumptions that date back several decades.
This interview with Paola Iacomussi examines where those legacy frameworks fall short for modern sensing stacks, how material variability and outdated reference observers affect perception reliability, and why a shift toward sensor focused standards is now essential for safe and scalable autonomy.
Written by:
Paola Iacomussi
Senior Researcher
Applied Metrology and Engineering Dept.
What assumptions do perception systems make about road markings and signage, and how old are those assumptions?
The road environment is designed to ensure the correct perception of all signages and of the whole environment by a driver. The assumption is that driving on a road is a task of sensing the road environment and acting accordingly, but who perceives the environment? obviously, a human being. Once upon a time it was a human only activity. Cameras and sensors (e.g., lidar) are now crucial to sensing and perceiving road surroundings, they compensate for human limitations and increase driving safety, but all road element performance is designed to satisfy human perception capabilities, not camera and sensor capabilities.
The case of road marking and signage perception is very significative: it is the case of an ineffective loop where vast perception potential is bent to human perception limitations. In road environments the semantic meaning of signage and road markings is conveyed by colours, that is a human perception uniqueness. Cameras and sensors must adapt their perception to material peculiarities, that are compliant to regulations based on the (limited) human reference observer.
The roots of this approach are historical: at the turn of the 1930s and 1940s, the first publications on the leading role of luminance and contrast in the road environment appeared, while in the 1960s-70s, the standards fixed the reference conditions for the characterisation of the whole road environment: materials (signage, road marking, asphalt), viewing conditions and road lighting conditions. Minimum performance requirements changed over the years thanks to technological evolution, but the conditions under which these are tested, are still those defined more than 50 years ago. The perception framework is different now.
Where do legacy road standards cause the biggest issues for modern sensing stacks?
Currently, road standards are predicated on the assumption that driving on a road is a visual task (of a human driver) of sensing the road environment and acting accordingly. All requirements refer to the human visual system in a given reference position with reference to the road environment, as was defined in the 40’s.
Driving on a road has transitioned from a human task to a technologically driven experience, where sensors are far away from being similar to a human observer for whom perception of all road materials and signage performance has been optimised. We are still linked to the legacy of the 40’s to assign requirements…
For example, if we consider the case of colours both at the level of road-marking and signals: colour conveys a semantic meaning for humans and we force cameras to go through colour recognition to reach the related semantic meaning. Wouldn’t it be more effective to go straight to the point of semantic meaning?
Do we need a biological reference observer for accurate perception, and, if not, how do we get around the HIL labelling?
If we want an accurate description of an observer’s perception, we must have a reference as close to the observer as possible. Whatever the observer is. Not necessarily and not for all applications does this reference always have to be representative of the human (observer, driver…). Currently, several databases on the spectral sensitivity of animal species and insects are available, for example we can find the reference bee observer and understand its perception of light (not limited to visible light), this is very interesting for applications with the environmental impact of artificial light.
If we extend this to the case of automotive sensors, we currently have good, very good knowledge of their performance, and we test them in an environment with material performance optimised for the human reference. So, for accurate perception of the road environment by sensors, we do not need a biological reference observer, but we need a reference representative of the “digital” non-biological observer.
We need normative references and regulations optimised both for human and sensor perception. For example, the colorimetric performance of a traffic light is to ensure the human observer’s correct perception of the colour and of the semantic meaning. In a sensor perspective, the semantic meaning STOP or GO should be perceived, because the route through the colour can be unnecessary long.
How does global variation in road materials affect perception reliability?
The heterogeneity of materials in the road environment is very wide. Some are highly heterogeneous by production, such as asphalt (aggregates), others change a lot over time and geography, as they are constantly exposed to weather and sunlight, that vary strongly with the latitude.
Let us consider the case of asphalt as an example: by oversimplification in simulation the road is a diffusing material (reflects incident radiation uniformly in space). But this is not the case: depending on the material and composition, it has a more pronounced directional behaviour (specular reflection) related to the size of the aggregates and the bitumen density. This directional behaviour increases or decrease over time. Not to mention the performance life of colour or retroreflectivity with time and exposure conditions. All these effects add up to an overall global variation of performance, then we add the uncertainty of sensor perception.
Describing the scene this way might suggest that nothing can be simulated accurately, or that everything is so variable that reliability is low. This is not the case. We have developed very detailed knowledge and measurement capabilities to ensure high reliability, but first we have to raise awareness that oversimplifying road material behaviour and radiation interaction affects the sensor’s perception reliability.
We are facing a step change similar to what happened with VFX effects in cinema. Do you remember the early special effects of the 90’s? They were based on oversimplified assumptions on the reflective behaviour of materials (diffusing and/or specular), then radiosity algorithms and accurate BRDF libraries of many materials arrived. The rest is history: and now it’s hard to perceive a VFX…
If road standards were redesigned for autonomy today, what would need to change first?
If we were to redesign road regulations, we would need to keep in mind the significant advances made by technologies. By technologies, I don’t just mean only automotive related like sensing, but also materials and lighting technologies. There is an urgent need for a new regulatory landscape to support technology and its efficient implementation, but also to acknowledge that the basic concepts on which regulation were based on 50 years ago are no longer valid. The premise that everything revolves around a human observer and their perception must be expanded to embrace materials technology and sensor needs and perspective.
We need to redesign standards and design approaches for the entire normative landscape of the road environment: from road surface materials to retroreflective materials and signalling. For example, LEDs used for both public lighting and signalling, have strong emission directionality and peculiarities which are partially neglected by current regulations.
Street lighting regulations provide tables describing the reflection coefficient of old asphalts and for geometries not relevant to LED lighting. Additionally, each value in the table is weighted by the cosine to the third power of the incident angle of light, because it would lighten computer calculations. How much computational time do we save now by not calculating the cubic cosine of an angle in an equation? What is the impact of considering light coming from useless directions? It is time to redesign.
Explore more on the AutoSens YouTube Channel⬇
