Emerging AI technologies are helping to create smarter traffic light timings to fight congestion and jams around special events, and beyond – creating the intersections and signals of tomorrow
Modern cities heave under the weight of peak-time traffic all year round, but during the balmier months of the year congestion levels are even higher. Chicago for instance registers a 1.5-hour increase in daily weekday congestion in June compared to January. Then, when 80,000 Taylor Swift fans descend on an urban stadium at the height of the season, the roads are forced to handle many times the capacity they were designed for. However, cutting-edge AI traffic management technologies can reduce the impact of such large-scale urban events.
Professor Mauro Vallati is the UKRI Future Leaders fellow in AI for Autonomic Urban Traffic Control (AI4UTC) at Huddersfield University. As Simplifai Systems, Vallati and the AI4UTC research team have been granted UK patent rights for their novel approach to traffic signal optimization in high-congestion contexts.
At its core, the system fine-tunes traffic signals based on historical data and changing conditions on the ground. It creates highly accurate, granular, simulations, which can then be used to generate signal optimization strategies as instructions for SCOOT.
Councilor Mark Ieronimo, portfolio holder for Transportation, Roads and Highways, Hull City Council
“The positive impact of using AI technologies to aid traffic management is evident… The council’s vision is to create a smart, connected city where mobility is efficient, safe and sustainable”
“Traffic signal optimization is the main means to control traffic directly,” says Vallati. “What we have designed works as a kind of plug-in on top of existing urban traffic control.”
Traffic operators specify one or more goals, and the system simulates traffic flow with different signal strategies until it finds the most effective approach.
“The goal can be to minimize the journey time on a corridor, allow as many vehicles as possible into the corridor, or push as many vehicles as possible out of the corridor, maximizing the throughflow,” explains Vallati.
It also receives information from SCOOT’s available sensors to get a read on current conditions and select an appropriate, pre-generated strategy.
“In AI, there are two main families, one is the machine learning, data-driven approach, and the other is model based,” Vallati explains. “In the former, you let the system learn everything from scratch, and it gives you a black box. You don’t know why or how it works, but it works.” AI4UTC has opted instead for the model-driven approach, which captures how real-life systems operate through explicit rules and representations.
Road testing
In June 2023, Vallati’s team had the opportunity to test the system in a real traffic situation when John Smith’s Stadium in Huddersfield hosted alt-rock band Muse. “We took charge of one corridor while the other remained under the usual traffic control system,” says Vallati. “We compared the results between the two corridors and with historical events, like a Green Day concert. According to that comparison, we saw a significant reduction in congestion. The number of vehicles queuing and delayed was reduced by 60%.”
The simulation’s predictions of its own efficacy even leaned on the conservative side compared to actual outcomes. “The strategy outperformed its own predictions by between 10 and 20%, so that was quite impressive,” says Vallati.
The sold-out concert coincided with peak traffic hours, and despite this, the AI-managed corridors were improved to a comfortable level. “The traffic was free flowing, while the rest of the corridors were completely stuck,” says Vallati. “This was also estimated to have a significant reduction in terms of pollution in the area.”
There were about 33,000 fans in attendance, and the traffic generated constituted about 10,000 extra journeys through the area, compared to ordinary peak-hour conditions.
Game changer
The test was so successful that Kirklees Council is now considering extending the system to the entire area around the stadium.
“There are of course challenges in terms of scalability because you need to have the capability to move the attention of the systems to where it is needed,” says Vallati.
Following successful trials, Hull City Council has awarded Simplifai Systems a two-year contract to develop a city-wide approach, with Councilor Mark Ieronimo, portfolio holder for Transportation, Roads and Highways calling AI traffic management integration a “game changer.”
“The positive impact of using AI technologies to aid traffic management is evident, with these figures showing the effectiveness of implementing these technologies,” says Ieronimo. “The council’s vision is to create a smart, connected city where mobility is efficient, safe and sustainable.”
“We envision a city where AI-driven traffic management becomes an integral part of our urban infrastructure,” adds Sean Higgins, ITS manager at Hull City Council. “The lessons learned from our initial trials will inform our decisions as we plan for further implementation of the system across Hull.”
Sensor connection
Sensors play an integral role in traffic management, but traditional sensors, such as inductive loops, are not fully optimized for rich data acquisition and deliver only basic information. However, VivaCity is one company that has designed an AI traffic monitoring system from the ground-up using its own computer vision sensor that can be installed at any roadside and is able to understand almost any traffic situation, using AI.
“Data needs to be highly accurate and truly multimodal to ensure control systems can respond to sudden peaks in demand”
Mark Nicholson, co-founder and CEO, VivaCity
“Data needs to be highly accurate and truly multimodal to ensure control systems can respond to sudden peaks in demand, including demand for road crossings close to venues from large numbers of pedestrians, and from new mobility such as e-scooters,” says Mark Nicholson, co-founder and CEO, VivaCity. “For example, at the AO Arena in Manchester, and close to football stadiums in Cambridge and Shrewsbury, VivaCity is helping to manage demand from all road users, but especially to balance pedestrian and vehicle demand at the end of events.”
More than 5,500 VivaCity sensors are currently in use globally. As sensors are always connected, anonymous data from the network of sensors can then be used to retrain the AI agent to provide further performance and accuracy improvements, benefiting cities and venues in making informed decisions about improving signal timings.
“In addition to real-time detection for signal control, sensors enable historical traffic monitoring and road safety analysis,” says Nicholson. “ New features and enhancements can be added over-the-air, making the sensors more future-proof than many other devices.”
The rich data from VivaCity sensors enables control systems to make better decisions. Its sensors are an integral part of the signaling and network management ecosystems, providing the ‘eyes’; with the ‘brain’ is provided by systems such as Simplifai and SCOOT.
Real-time ambition
At this stage, the Simplifai system cannot operate in real time, but pre-generates simulations and outputs strategies, to be chosen depending on traffic conditions as revealed by sensor input.
“The simulation is capable of running in real time, but delays in releasing sensor data precludes this”
Professor Mauro Vallati, UKRI Future Leaders, AI4UTC, Huddersfield University
“It’s not a limitation of our system,” says Vallati. “The entire simulation is capable of running in real time, but delays in releasing the data from available sensors precludes this. Currently we leverage existing sensors, but the data takes a few hours to be released by the SCOOT system,” says Vallati.
A further limitation is that strategies must be pre-uploaded to the existing system. “For the Muse trial we ran the simulation the day before and uploaded the settings that we expected to be used,” says Vallati. “You are telling the traffic controller in charge to switch to a different pre-generated strategy. So that is reducing the possibilities and the freedom that you have. Ideally you would have access to the data from sensors in real time, allowing you to alter junction settings straight away. Assuming that they have gone through validation, so that they are correct and they are safe.”
Given the impressive results of early trials, Vallati is confident that AI4UTC’s system can aid many municipalities in their congestion issues, for large events and everyday traffic.
Huddersfield’s John Smith’s Stadium. Photo: Arne Müseler
Machine learning
AI is also helping control signals in the UK city of Coventry – a reinforcement learning technique is proving successful in a unique trial on three junctions. Here, it wasn’t a government highways body, or a private firm seeking a lucrative opportunity that led to the development of this unique CCTV-driven AI traffic management system, it was one individual’s frustration over traffic jams.
Dr Maria Chli, from the Computer Science Research Group at the Aston Centre for Artificial Intelligence Research and Application had just had enough of the A38 in Birmingham. “It started out as personal interest, being stuck in traffic and feeling that we can do things better,” she recalls. “Depending on the time of day, there is a tidal wave route to get in and out of Birmingham, controlled by a human. It just didn’t work.”
Initially starting in 2018, two simulations were formulated, one a bit like a computer game and a more successful version generated by what CCTV observes. This led to work with traffic lights and to a full trial currently underway in Coventry.
“Reinforcement learning is a technique where the machine learns the job by receiving rewards,” explains Chli. “The AI senses the environment and it has some actions at its disposal. It learns, over time, which action to take to achieve its goals. It starts by exploring its portfolio of actions totally randomly, and associates those actions with certain desirable environmental states. We have taken the next step to use deep reinforcement learning, where we don’t prescribe what the environment should be, we just offer the system the CCTV pictures and it learns to associate certain pixels with their rewards.”
Dr George Vogiatzis, a reader in Computer Science at Loughborough University helped to design the system. He explains, “It’s essentially a trial-and-error approach, dressed up with mathematics. The AI responds to certain signals, and the better the reward the more it knows it is getting it right.”
Chaos could ensue if this is trial-and-error process were to be put straight on the road, so the training simulator stage is crucial. The current one in use is indistinguishable from CCTV footage, and the goal here is to diversify the inputs the AI receives. This includes a variety of weather conditions and different levels of darkness and light, as well as the traffic actors and what’s around them.
“This is important because of the sim-to-real gap,” Chli notes. “In other words, AI trained on simulation and then unleashed on the real world will experience a gap in its understanding and ability to act. To minimize this we need to give it as many different experiences as we can before it gets out there.”
This training process doesn’t take long. After some experimentation, the pair have settled on three days as an ideal timeframe. “You have to be careful not to overtrain it, in case it forgets what it learnt first,” says Vogiatzis. “Also, it continues to learn on-the-job.”
Real-world deployment
Simulations complete, the first systems went live on three junctions in Coventry in November 2023. These are initially being looked at in isolation, but then as part of a network, and Chli is conscious of the risk of moving a traffic problem elsewhere. The philosophy is not to try and tackle the complexity of networking everything, but
to concentrate network efforts where the rewards will be the greatest. Using AI is particularly helpful here.
“With parameters such as flow, throughput or waiting times we’re doing particularly well, especially with dense or more unpredictable traffic,” says Chli. “We’re at a crucial point now with good data emerging.”
AI classification of vehicle types by VivaCity
In the future, environmental information could also be added to the model to refine it. “We know that air quality improves as waiting time is minimized at the lights,” says Chli. “Also, air-quality worsens when vehicles are braking, as tires and discs shed particulates so we may not want to stop or slow down cars unnecessarily. It’s all complicated by the effect that tall buildings have, trapping the particulates.”
Here Vogiatzis sites UK government figures, attributing up to 15% of CO2 emissions to inefficient traffic management. “Projects like this will be crucial when it comes to achieving net zero and also general improvements in urban air quality,” he says.
Priority please
An important part of the design of the system is the ability to isolate certain vehicles, to give them priority. “We use the reward signal again for this, in a similar way to training a dog,” says Vogiatzis. “Like giving the dog treats for being good, we reward the AI for prioritizing certain vehicles that it recognizes, such as ambulances, and allowing them to pass through quickly.”
Just as the dog may not comprehend good or bad behavior, it understands it will be rewarded for certain things, so modulating rewards allows the system to be fine-tuned. So, for example, emergency vehicles could get more rewards than buses, which get more rewards than cars. Another feature of this is the ability to specify what is prioritized. For example, at a junction near a school – or when a special event is scheduled such as a rock concert – pedestrians could be priority, but only when the event or school day starts and finishes.
It’s good but not fool-proof. Vogiatzis recalls some of the early research in London when the system struggled to differentiate between buses and Royal Mail trucks of a similar size and color, leading to both being given priority.
“We must learn from this, especially as we are now on the cusp of giving AI more authority. Bad agents may attempt to hijack and take advantage of AI, so the design and testing of these new systems must be very robust. The positive side of this is that we don’t have to do any vehicle tracking or ALPR, practices which we know are unpopular with the public. It’s a very passive system. It doesn’t need to store information.”
The way forward
Another potential advantage of employing the system is route planning to avoid congestion in urban areas, particularly at peak-traffic times, such as during special events.
“Could it be in the future that we have traffic lights that tell us which way to go? Could it be that sat navs or Google Maps send traffic different ways in order to create a more uniform flow?”
Dr Maria Chli, Computer Science Research Group, Aston Centre for Artificial Intelligence Research and Application, UK
“Could it be in the future that we have traffic lights that tell us which way to go?” asks Chli. “Could it be that sat navs or Google Maps send traffic different ways in order to create a more uniform flow? This is something we’re studying. It has the added difficulty of drivers not complying, so the key here is to offer to follow the routes we are proposing. This concept will emerge in the next few years.”
Looking further into the future, both see the potential for widespread adoption, not least because of the very low cost. With conventional traffic management systems costing around £150,000 (US$188,000) per junction, and a good camera system as low as £3,000 (US$3,800), the economics are clear.
Despite the growth of vehicle autonomy, the continued existence of human drivers means some degree of external control will still be needed. Chli concludes, “We will still have traffic lights for many decades to come, this is just a better way of managing them.”
The benefits of AI are still being explored. It’s a good time for event organizers and traffic management professionals to become ‘AI-literate’ and understand the various advantages and disadvantages of current and future offers.
