Fraunhofer Project: Intelligent Traffic Lights for Better Traffic Flow

Previous traffic light control systems operate based on rules and are equipped with induction loops embedded in the asphalt. However, these rules and existing sensors only roughly represent traffic situations, according to a research team at the Department of Industrial Automation INA of the Fraunhofer Institute for Optronics, System Technologies, and Image Exploitation IOSB in Lemgo.

As part of the KI4LSA project, they are researching how an AI-based traffic light control system can proactively manage traffic. High-resolution camera and radar sensors are also intended to capture traffic precisely. The goal: fewer congested intersections, better traffic flow, and more safety, especially for pedestrians.

Real-time vehicle count and lane-accurate

The number of waiting vehicles at the intersection is to be recorded in real-time and lane-accurately. Also to be recorded are the average speed of the vehicles and the waiting time. Real-time sensors are combined with artificial intelligence. The AI uses algorithms from so-called Deep Reinforcement Learning - a method of machine learning that aims to find intelligent solutions for complex control problems.

“We have built a realistic simulation model of the Lemgo intersection where our tests are taking place, and trained the AI in this model with countless iterations. Beforehand, we transferred the measured traffic volume during rush hour into the simulation model so that the AI can work with real data,” explains Arthur Müller, project manager and scientist at Fraunhofer IOSB-INA.

A trained agent, a neural network, represents the traffic light control system, and the algorithms trained in this way are supposed to determine the best traffic light switching behavior and the best phase sequence. This is expected to reduce waiting times at the intersection, driving times, noise created by traffic jams, and CO2 emissions.

Up to 15 percent better traffic flow

The AI algorithms run on an edge computer in the control box at the intersection, the researchers said. They can be tested, applied, and scaled on networked control systems, such as adjacent traffic lights that are part of a network. Simulations at the test intersection in Lemgo, which was equipped with such intelligent traffic lights, showed, according to the research team, that traffic flow could be improved by 10 to 15 percent through artificial intelligence. Further evaluations are planned for the coming months, with the trained agent to be brought onto the street, i.e., transferred to the real-world laboratory. Parameters such as noise pollution and exhaust emissions are also to be taken into account. However, certain hurdles must be overcome:

“The assumptions about traffic behavior in the simulation do not match reality one-to-one. Accordingly, the agent needs to be adjusted,” says Müller. “If this succeeds, the scaling effect is enormous, considering the large number of traffic lights in a city like Lemgo alone.”

According to Müller, AI-controlled traffic lights could make better use of the existing infrastructure. According to the press release from Fraunhofer IOSB, traffic jams would cause an annual economic damage in the EU, estimated at 100 billion euros for the member states.

"Worldwide, we are the first to test traffic light control through Deep Reinforcement Learning under real conditions. We are banking on the exemplary character of our project,” explains Müller.

The project runs until the summer of 2022 and is funded by the Federal Ministry of Transport and Digital Infrastructure (BMVI). Project partners include Stührenberg GmbH, Cichon Automatisierungstechnik GmbH, Stadtwerke Lemgo GmbH, as well as the Old Hanseatic City of Lemgo (associated) and Straßen.NRW (associated).

Pedestrians in focus: traffic light based on demand

Another project named KI4PED focuses on pedestrians: Together with Stührenberg GmbH and the associated partners Straßen.NRW, the city of Lemgo, and the city of Bielefeld, Fraunhofer IOSB-INA aims to develop an innovative approach to pedestrian traffic light control based on demand by the end of July 2022. Particularly in focus are vulnerable persons, such as the elderly or people with disabilities. The goal is to shorten wait times and increase safety at traffic light intersections through longer crossing times because the green phases for these groups are currently too short, according to recent studies.

Currently installed push buttons provide too little information, such as the number of people, the researchers say. Using artificial intelligence and high-resolution lidar sensors, crossing times are to be automatically adjusted to the needs of individual pedestrians. Person recognition and tracking will be achieved based on lidar data using AI and implemented in an embedded system in real-time.

“For reasons of data protection, we use lidar sensors instead of camera-based systems, as they represent pedestrians as 3D point clouds, which cannot be identified,” explains Dr. Dennis Sprute, project manager and scientist at Fraunhofer IOSB-INA.

Lidar sensors emit laser beams to measure distance and register the reflected light. The light travel time provides information about the distance to the object, i.e., the person. These sensors are also robust to lighting, reflection, and weather conditions.

A feasibility study examines their optimal positioning and alignment at the traffic light intersection. The AI algorithms will be trained for one week at two traffic light intersections in Lemgo and Bielefeld. Sensor tests are also planned on the grounds of Fraunhofer IOSB-INA under various simulated lighting conditions to determine recognition performance.

According to their statements, the research partners hope that with a demand- and situation-based control concept, wait times during high pedestrian traffic can be reduced by 30 percent and the number of dangerous illegal crossings can be reduced by about 25 percent.