150x150.png)
Coen Bresser, Senior Manager for Innovation and Deployment at ERTICO – ITS Europe, talks with Alistair Gollop about his involvement in navigating the emerging complex landscape of Artificial Intelligence (AI) in traffic systems.
Artificial Intelligence (AI) is no longer a futuristic buzzword in transport; it's increasingly embedded in the systems managing our road networks. From optimizing signal timings to predicting congestion hotspots, its potential seems boundless. Yet, harnessing this power effectively presents significant challenges, requiring unprecedented collaboration, robust safety protocols and a clear-eyed view of both its capabilities and limitations.
A fundamental challenge arises from the fragmented nature of current traffic information. While navigation apps offer invaluable real-time guidance, their independent operation can inadvertently worsen congestion. "There are various examples of what can go wrong," Bresser notes. "If there's a traffic jam on the highway and you use a [navigation] app... all the people who are diverted end up in the same school yard, for example. So, you just shift the congestion somewhere else."
This phenomenon, sometimes echoing the principles of the Braess paradox (where adding capacity can paradoxically decrease overall efficiency), highlights the need for coordination. The TM 2.0 initiative aims to bridge this gap. "What TM 2.0 aims for is to create cooperation," explains Bresser, who co-chairs the platform. "By sharing the intention of traffic management towards the participants driving there... they can take into account what will happen."
This involves not just one-way communication from traffic managers to a single service provider, but collaboration between service providers (like Google, Apple, Waze, TomTom, HERE) and public authorities. If competing services optimize routes independently, they might funnel vehicles onto the same alternative routes, creating new bottlenecks. Cooperative strategies, guided by public authorities' overarching goals (like maximizing network flow or prioritizing certain vehicle types), could enable traffic to be spread more logically across the entire network, minimizing overall disruption.
This introduces the concept of "co-opetition" – a blend of cooperation and competition. "Service providers have their own customer base... or they offer services that their competitors don't," Bresser acknowledges. "But still, you can cooperate on data sharing, for example. If you share data with someone else and you get it back, you both have a better view of what's going on." Listening to traffic management directives, such as planned road closures and suggested alternative routes, ultimately enhances the value proposition of cooperating service providers. It counters the fear that incorporating external public authority choices might worsen their individual routing algorithms. As Bresser points out, "This road will be closed anyway," and failing to account for it damages user trust – "Why didn't you know this?"
This trust is fragile. Conflicting messages – a matrix sign instructing drivers left while their app says right – cause confusion, uncertainty and potentially erratic driving behaviour, ultimately reducing safety. Harmonizing information through collaboration ensures drivers receive consistent, reliable guidance, reinforcing the credibility of both public infrastructure and private services. Drivers, pedestrians and cyclists often only perceive their immediate surroundings; conveying the "bigger picture" benefit of coordinated traffic management is crucial but challenging.
While generative AI captures headlines, Bresser emphasizes that AI is not new to traffic management. "AI has been used for over two decades in traffic management," he states. Traditional AI and machine learning have long been employed for tasks like object detection via cameras, traffic flow modelling and incident prediction.
The game-changer lies in AI's ability to handle escalating complexity. Modern traffic management systems offer an ever-increasing number of options (signal timings, variable speed limits, ramp metering). Manually managing these, especially through pre-defined "scenario-based" responses, becomes exponentially harder. "With these advanced traffic management systems, the number of scenarios explodes... to such an extent that there is a limit that you cannot comprehend the entire set," says Bresser. This is where AI, particularly newer forms, can play a vital role. "AI might be able to assess, given this set of scenarios in this traffic situation, 'I would advise this set of settings'," Bresser suggests. AI can act as a powerful decision-support tool, sifting through vast amounts of data and potential outcomes far faster than a human operator, presenting optimised strategies for consideration.
Despite AI's analytical prowess, ERTICO and Bresser strongly advocate against allowing AI autonomous control over safety-critical systems. "We particularly state that AI that is allowed to freely operate in the functional space – so everything on street – you cannot control that... we should not want that," he stresses. The risks are too high. AI systems, fundamentally statistical machines predicting probable outcomes ("if this, then probably this is what you mean"), can make mistakes. Allowing an AI to independently control traffic signals without hardwired safety interlocks could, theoretically, lead to catastrophic failures like giving green lights to conflicting approaches simultaneously. "In Europe, it's physically wired not possible," Bresser reassures, highlighting the existing safeguards. "To safely operate AI in traffic management, these safeguards are a prerequisite." A prime example of responsible AI implementation is hard shoulder running. AI-powered object detection systems monitor camera feeds far more efficiently than human operators, automatically closing a hard shoulder lane if an obstruction (like a broken-down vehicle) is detected, significantly enhancing safety. However, the decision to re-open the lane remains firmly with a human operator.
Beyond operational safety, AI deployment raises profound ethical questions. Training data bias is a major concern. Models trained predominantly on data from one demographic might be less effective at detecting individuals from other groups, creating inherent safety inequities. Ensuring diverse and representative training datasets is crucial, as is transparency about how models are trained and validated. The "trolley problem" – hypothetical dilemmas about unavoidable accidents – often surfaces in discussions about autonomous systems. While extreme, it highlights the need for societal consensus on decision-making logic embedded in AI. Bresser offers a pragmatic perspective: "If I'm driving... and I get to choose if I kill myself or the other, lucky me will do my utmost best not to kill the other. But of course, hardwired in your brain is yourself." He argues that the focus should be on whether AI improves safety compared to the human baseline, even if imperfectly. Refusing deployment because of edge-case ethical dilemmas or biases might mean foregoing significant overall safety gains. Establishing clear probability or certainty levels for AI recommendations could aid transparency and trust.
A persistent challenge is demonstrating the value of ITS interventions, especially those enhancing safety. "The value lies in the absence of the transaction, so being the accident," Bresser observes wryly. "How are you going to monetize something that's not there?" Unlike a consumer product, safety systems prevent negative outcomes. Proving their effectiveness often relies on analysing trends in accident reduction over time, making direct ROI calculation difficult. This necessitates investment driven by societal value, typically led by governments, public authorities, or entities like insurance companies that benefit from reduced incident costs across large populations.
With the foundational concepts gaining traction, the next critical phase for TM 2.0 and AI in traffic management is deployment. "What we need is to move towards the deployment of the concept," Bresser asserts. This involves clarifying the steps needed for stakeholders – road authorities, service providers, technology suppliers – to participate. "We need to clarify what the tools are, what the products are, what the standards are that you should use," he outlines. The goal is to create a practical roadmap or manual enabling organizations to understand how they can contribute to and benefit from this collaborative, AI-assisted future of traffic management. Guiding authorities, often risk-averse due to budget constraints and fear of making incorrect investments, is paramount. AI offers transformative potential for ITS, promising smoother traffic flow, enhanced safety and better-informed travellers. However, unlocking this potential requires moving beyond isolated systems towards genuine co-opetition, embedding robust safeguards, addressing ethical considerations head-on and providing clear pathways for adoption. The journey, guided by collaborative platforms like TM 2.0, is complex but essential for navigating the future of mobility.
Watch the whole interview on the @ITSNow YouTube Channel -
See the video