“Roads are chronically congested and vehicles queue endlessly at junctions. Rush hour is especially bad for long traffic jams. At the Fraunhofer Institute for Optronics, System Technologies and Image Exploitation IOSB, researchers in the institute branch for industrial automation INA in Lemgo are using artificial intelligence for smart traffic light control as part of the “KI4LSA” and “KI4PED” projects. In the future, self-learning algorithms combined with new sensors should ensure better traffic flow and shorter waiting times, while providing improved safety for pedestrians at crossings…Instead of conventional sensors, they are using high-resolution cameras and radar sensors to more precisely capture the actual traffic situation. This allows the number of vehicles waiting at a junction to be determined accurately in real time. The technology also detects the average speed of the cars and the waiting times. The real-time sensors are combined with artificial intelligence, which replaces the usual rigid control rules. The AI uses deep reinforcement learning (DRL) algorithms, a method of machine learning that focuses on finding intelligent solutions to complex control problems. “We used a junction in Lemgo, where our testing is carried out, to build a realistic simulation and trained the AI on countless iterations within this model. Prior to running the simulation, we added the traffic volume measured during rush hour into the model, enabling the AI to work with real data. This resulted in an agent trained using deep reinforcement learning: a neural network that represents the lights control,” Arthur Müller, project manager and scientist at the Fraunhofer IOSB-INA, explains the DRL approach. The algorithms trained in this way calculate the optimum switching behavior for the traffic lights and the best phase sequence to shorten waiting times at the junction, reduce journey times and thus lower the noise and CO₂ pollution caused by queuing traffic. The AI algorithms run in an edge computer in the control box at the junction. One advantage of the algorithms is that they can be tested, used and scaled up to include neighboring lights that form a wider network…”