Traffic Signals That Learn
Threading a line of orange barrels at midnight, with the highway lane pinching down and a semi filling your mirror, is nobody's idea of a good time. It is also more dangerous than it feels. Work zone crashes killed 850 people on U.S. roads in 2024.
That was down from 905 the year before, but it still works out to more than two deaths every day. Almost half of fatal work zone crashes happen after dark, according to the Federal Highway Administration, the hours when traffic thins out and speeds climb. Speeding is a significant factor in more than 30% of fatal work zone crashes, and at night the only thing standing between a work crew and a moving truck is a row of plastic barrels and a temporary sign.
The most common way these accidents happen is also the most ordinary. Rear-end collisions accounted for 21% of fatal crashes in 2022. Work zone crashes also cost the country about $38.9 billion a year in comprehensive societal costs.
Georgia sits in the middle of these issues. The state runs more than 500 active work zones at any given time, and it is ranked fourth in the nation for fatal work zone crashes involving large trucks.
Agencies move construction to overnight hours to spare daytime drivers the backups. That trade solves one problem and creates another: it moves the risk into hours when work zones are at their most dangerous.
A new study published in Future Transportation asks whether artificial intelligence can help traffic signals respond to both work zone problems at once: reducing traffic congestion while also reducing safety risks around the work zone merge. The study was led by researchers with expertise in transportation engineering and roadway safety from various organizations, including Parsons Corporation, Jacobs Engineering, Iowa State University, and The University of Alabama.
The researchers developed a reinforcement-learning traffic signal controller for work zones: a signal that adjusts its own timing as the traffic in front of it changes. Rather than following a fixed plan, it watches what is approaching, decides how long to hold each movement and continuously updates that decision throughout the night.
A signal that reads the road, not a script
Part of what makes overnight work zones dangerous is that the traffic signals around them are not watching anything. They are reading a script. When a highway crew closes a lane, the temporary signal timing that goes with it is usually worked out in advance based on assumed traffic volumes and then left alone for the duration of the job. That is a reasonable way to handle a road whose traffic looks the same every night. Overnight traffic doesn't. A group of trucks comes through at 1 AM, and the queue jumps; the next 40 minutes are nearly empty. The signal does not know the difference. It runs the same plan for both.
In the Future Transportation simulation, the AI-driven reinforcement-learning traffic signal controller moved considerably more traffic through the work zone than the conventional fixed-time signal plan: roughly 25% more cars, about a third more trucks and nearly 40% more buses. The longest queues shrank by about 40%. The AI-driven controller also produced the fewest risky merges and the least stop-and-go crawling of any of the controllers tested.
The more interesting result is what did not happen. Similar systems were built to move vehicles with less travel time, less idling and better throughput. This one is scored on safety and penalized when queues back up too far, when traffic falls into repeated stop-and-go cycles, and when merging near the lane closure turns unstable. The expectation going in was that safety would cost something in throughput. It did not. Calming the merge and keeping the queue from backing up protected the work zone's capacity, so the safer plan was also the faster one.
Measuring danger without waiting for crashes
Measuring whether the signals made the work zone safer required a stand-in for crashes, since you cannot wait for real ones.
The simulation tracks the conditions that make a merge go wrong: how close vehicles are, how fast they are closing on one another, and how sharply drivers accelerate and brake while hunting for a gap near the lane closure. It watches queues building upstream, and the repeated stopping and starting that comes with them. These are standard surrogate measures in traffic safety research. They identify dangerous conditions; they do not predict or count crashes.
The next step is to get it off the computer. The research team is working with the Georgia Department of Transportation to move the approach toward a field test, possibly in Midtown Atlanta, with the goal of eventually running it in active work zones and seeing how it holds up when the traffic is real.
Simulated drivers are patient and predictable in ways that drivers at one in the morning are not. They do not tailgate out of impatience, misjudge a gap, or look up from a phone to find the lane gone. A controller that performs well in a simulation world has shown it can work in principle. A live work zone is where it finds out.
There is a second reason to want it on a real road. The controller's whole advantage is that it sees a queue forming before the drivers heading into it can. Right now, the only way it can pass that on is by changing a light. But the same information, that there is stopped 600 feet ahead and the lane ends just past it, is exactly what a connected vehicle could receive directly, and what an automated one would need. Work zones are among the hardest places for a self-driving system to read: temporary striping, cones that moved since the last map update, a lane that ends where the road is supposed to continue. A signal that already knows what the traffic ahead is doing is a useful thing for such a car to ask.
Studying danger before facing it
Another project extends the same safety-first approach to self-driving vehicles, creating rare but realistic dangerous scenarios, such as a pedestrian emerging between parked cars or a cyclist appearing in a vehicle’s blind spot, so automated driving systems can be tested before encountering those situations on real roads. The work expands an NVIDIA-developed framework to include pedestrians, cyclists and motorcyclists, addressing scenarios that are both difficult to capture in real-world datasets and too dangerous to reproduce experimentally.
Both projects follow the same principle: dangerous situations should be studied in simulation before people are exposed to them. Using Georgia Tech’s PACE computing cluster, the researchers can repeatedly train and test these systems across thousands of simulated scenarios.
The technology is not intended to replace signs, barrels, speed limits, or other established safety measures. It adds something different: attention.
Whether it is a traffic signal recognizing a dangerous queue before a crash occurs or an automated vehicle prepared for a rare pedestrian encounter, the goal is to build transportation systems that can recognize risk earlier and respond before that risk becomes a crash.
Author Bio
Israel Afriyie is a transportation engineer and researcher based in Metro Atlanta. He works at Parsons Corporation and is a co-author of the Future Transportation study described in this article. His research focuses on using artificial intelligence and emerging technologies to improve traffic safety and mobility, including AI-driven traffic signals, autonomous vehicles and transportation simulation. He is also an active researcher and reviewer for engineering and transportation publications and is committed to developing practical solutions for safer and smarter transportation systems.