Three traffic lights against blue sky background
Traffic lights are for cars. But what if they could see waiting pedestrians and even prioritise them, without having to tap a “beg button” and hope it actually does something?
That’s one of many benefits promised by a project at Aston University, which pairs a low-cost 360-degree camera with AI analysis in order to better manage traffic — and the researchers have been running it on a live intersection in Coventry, UK.
“The AI makes a decision about each of the allowable combinations,” Dr Maria Chli of the Aston Centre of Artificial Intelligence Research and Application, explained on the sidelines of the Turing Institute’s AI UK conference this week.
AI traffic lights
Chli’s system uses deep reinforcement learning, so each decision is assessed against pre-decided metrics such as wait times. The team has developed an equation that decides if the AI gets a “reward” or “punishment” based on those metrics, helping to train it.
“Usually the first 100 decisions are rubbish,” she says. But once trained using simulations, it quickly learns how to manage traffic and is ready to go on live roads within days.
“When I train an AI in simulation, which is how I start, it takes about three days for a particular junction,” she explains. “Transferring previous knowledge, it’s actually pretty good.”
That said, it can be best to let such systems learn on the job, she said, as there’s a risk to overlearning — get too specific, and it does less well on situations it hasn’t seen before.
Smarter traffic lights – and cheaper too
Using a camera and a controller worth a few hundred pounds has a few benefits over existing systems, notably the magnetic induction loops that rely on a metal cable embedded in the road.
Those systems are expensive, require ripping up the road to install causing disruption, while this AI powered traffic management system achieves better performance at less cost, Chli says.
Plus, those cables embedded in the asphalt ignore other road users. And there’s more to traffic than cars — what about bikes and pedestrians? This system can include such travellers in its equations, and it can also prioritise them.
For example, the Coventry junction that Chli’s system was trialled at is next to a major hospital. The smart camera can not only ensure emergency vehicles get priority, but also the waves of staff trying to cross the road at shift changes at 8am and again at 8pm. “With this technique, we can tailor the priorities for each junction,” she says.
Smarter traffic lights for safer cities
Priority could also be given to cyclists or buses to help encourage people out of cars and in favour of active or public transport. Near schools, the system could offer longer crossing times ahead of start and finish times for classes, to help keep children safer and give them more time to cross the road.
“For me, aside from the cost, the biggest strength is that you can tailor it to junctions to help vulnerable road users without impacting privacy,” she said, as the system doesn’t identify specific people, just that they’re pedestrians or on a bicycle.
Of course, she notes, making a change at one junction impacts the rest of the system, so the next step for the researchers is to roll out the smart traffic management cameras across a wider network of multiple junctions.
The deep reinforcement system will work to get cars through quicker, but doing that while also including people in the equation is perhaps the real breakthrough.
