A new artificial intelligence system, developed by researchers at Aston University in Birmingham, UK is providing a new, hardware-light way to monitor traffic flows, and optimise signal phase and timing, using just video feeds.
It uses deep reinforcement learning, where a program understands when it is not doing well and tries a different course of action, or continues to improve when it makes progress.
In testing, the system significantly outperformed all other methods, which typically rely on manually-designed phase transitions.
The researchers built a state of the art photo-realistic traffic simulator, Traffic 3D, to train their program, teaching it to handle different traffic and weather scenarios. When the system was tested on a real junction, it adapted to real traffic intersections despite being trained entirely on simulations.
“We have set this up as a traffic control game,” explains Dr Maria Chli, reader in Computer Science at Aston University. “The program gets a ‘reward’ when it gets a car through a junction. Every time a car has to wait or there’s a jam, there’s a negative reward. There’s actually no input from us; we simply control the reward system.”
The main form of traffic light automation currently used at junctions depends on magnetic induction loops; a wire sits on the road and registers cars passing over it. The program counts that and then reacts to the data. Because the AI created by the Aston team ‘sees’ high traffic volume before the cars have gone through the lights and makes its decision then, it is more responsive and can react more quickly.
“The reason we have based this program on learned behaviours is so that it can understand situations it hasn’t explicitly experienced before,” says Dr George Vogiatzis, senior lecturer in Computer Science at Aston University. “We’ve tested this with a physical obstacle that is causing congestion, rather than traffic light phasing, and the system still did well. As long as there is a causal link, the computer will ultimately figure out what that link is. It’s an intensely powerful system.”
The program can be set up to view any traffic junction – real or simulated – and will start learning autonomously. The reward system can be manipulated, for example to encourage the program to let emergency vehicles through quickly. But the program always teaches itself, rather than being programmed with specific instructions.
The researchers hope to begin testing their system on real roads this year.
The research paper, Fully-Autonomous, Vision-based Traffic Signal Control: from Simulation to Reality by Deepeka Garg, Maria Chli, George Vogiatzis, is being presented at the Autonomous Agents and Multi-agent Systems Conference 2022 being held virtually May 9-13, 2022.
Images: Aston University.