Riding the “Green Wave”
Author: Markus Strehlitz
Not only can traffic lights be a nuisance, but they can also obstruct traffic if not optimally controlled. That is why scientists want to use artificial intelligence to adapt traffic light control to the current traffic situation at all times.
Controlling traffic in an intelligent way is one of the most pressing challenges when it comes to the mobility of the future. Traffic lights play an important role here. If their signal changes are optimally adapted to existing conditions, waiting times can be reduced – which not only contributes to a better flow of traffic, but is also easier on drivers’ nerves.
However, traffic light control systems currently operate according to rigid rules. And the sensors embedded in the ground that are used for this purpose only roughly map the traffic situation. This makes it virtually impossible to flexibly adapt the controls to existing traffic conditions.
Scientists are working to change this – with the help of artificial intelligence (AI). To do so, researchers at Aston University in the USA have developed a system that uses cameras and a simulation program. The AI software is first trained with a traffic simulator in order to learn how to deal with different scenarios. Integrated into a traffic light circuit, the adaptive system then accesses the cameras installed on site to analyze the images of the current traffic volume. The scientists tested the technology at a real intersection. The system was able to adapt to real-life situations and control the traffic lights accordingly.
To train the software, the researchers use deep reinforcement learning, which works more or less like a reward system. Every time a car drives through the intersection, the AI receives positive feedback. If a vehicle has to wait or if a traffic jam develops, it receives negative feedback. In this way, the AI software looks for its own way to optimally control the traffic lights. In the process, the reward system can also be set so that emergency vehicles, for example, are routed through faster. The scientists trained the system in this way so that it can respond to situations it has not faced before, explains Dr. George Vogiatzis, a lecturer in computer science at Aston University.
The goal is optimal traffic flow
Reinforcement learning is also used by researchers at the Fraunhofer Institute of Optronics, System Technologies and Image Exploitation (IOSB). In the KI4LSA project, scientists at the Institute for Industrial Automation INA in Lemgo have also developed a system to control traffic lights intelligently and predictively. The solution uses cameras as well as radar sensor technology and has been trained similarly to the system from Aston. “We built a realistic simulation model of the Lemgo intersection where our tests are taking place and let the AI train countless iterations in this model,” reports project manager Arthur Müller. “Before that, we transferred the traffic volume measured at rush hour into the simulation model so that the AI can work with real data.” The AI algorithms run on an edge computer located in the intersection’s control box.
According to Fraunhofer IOSB, AI could improve traffic flow by ten to 15 percent. This is the conclusion reached by the simulation phases at the congested Lemgo intersection, which was equipped with intelligent traffic lights.
But the technology is not yet in everyday use. “The biggest challenge in moving the system from a proof of concept to a product is to generalize our solution, apply it to different types of intersections, and integrate it into a grouped control of multiple intersections,” Müller says. Another challenge, he says, is to stabilize the AI’s training process.
Dr. Thomas Wagner, Head of DEKRA’s Assessment Center for Driving Suitability, is fundamentally concerned about the use of cameras to control traffic lights more intelligently. Data protection would stand in the way of such a concept in Germany. “Something like this is certainly more feasible in the USA than here,” says Wagner.
According to Müller, however, the data protection issue has definitely been incorporated into the project. The cameras are located at a height of seven meters with a vertical tilt angle of 40°. The resolution is only 640 x 480 pixels. “This is sufficient for classifying and determining the position of vehicles, but does not allow any personal reference – i.e., the identification of faces,” Müller explains. License plates are also generally not legible. “Furthermore, the material is processed directly on site and without human intervention; there is no storage or transmission.”
AI also helps pedestrians
Data protection also plays an important role in another project, in which the Fraunhofer IOSB also wants to help pedestrians by using AI. In the KI4PED project, the scientists are working with partners on an approach to demand-driven control of pedestrian traffic lights. The goal is to shorten waiting times and increase the safety, particularly of elderly and physically impaired people, by extending the window of time for crossing streets. “For reasons of data protection, we use lidar sensors instead of camera-based systems, as they display pedestrians as 3D point clouds and thus cannot identify them,” explains project leader Dr. Dennis Sprute.
The people responsible also expect significant improvements in this project. A demand- and situation-oriented control concept should be able to reduce the waiting time during high volumes of people by 30 percent and the number of dangerous crossings in violation of traffic regulations by about 25 percent.
Three questions for Dr. Thomas Wagner, traffic psychologist and Head of DEKRA’s Assessment Center for Driving Suitability
Mr. Wagner, do red traffic lights make people aggressive?
There are studies that deal with aggression in road traffic in a broader sense. They come to the conclusion that it is the interplay of several factors that leads to anger and annoyance among drivers. Whenever there is a discrepancy between desired speed and the speed that is actually possible, drivers become stressed. I.e. when the flow of traffic is severely impeded. Red traffic lights contribute to this, but so do many other conditions. The crucial point here is plausibility. If the driver does not understand why he has to wait so long at a traffic light or why a construction site is impeding his progress, his annoyance increases.
Does this also increase the risk of accidents? Can we say that the more stressed a driver is, the more likely they are to collide with another road user?
One danger of stress is that I become inattentive. I no longer pay sufficient attention to what is happening around me. The other risk factor is that when I’m emotionally on high-alert – that is, focused on fight or flight – a small spark is enough to make me explode. A small provocation is enough for me to disregard the rules and run a red light, for example.
In your view, is there a fundamental need to optimize traffic light controls?
That is certainly the case. And if you take a look around, some excellent solutions already exist abroad. For example, I’m talking about systems that show how many seconds a traffic light has left on red. Because one problem for drivers is that they don’t know how long they have to wait at a traffic light. Appropriate information would therefore take a lot of pressure out of the situation. Longer yellow phases or a flashing light before the traffic light starts again would also help to defuse the situation.