The city is a living organism, its streets veins through which countless vehicles, pedestrians, and buses move in a rhythm that never quite repeats. Researchers are getting closer to giving traffic signals a form of perception—so they can respond not just to a fixed script, but to the actual beat of daily life. A recent study from Marian University and the University of Indianapolis shows how we might teach traffic signals to listen to that beat by turning cameras into data, and data into smarter light cycles. The work is led by Mohammad Shokrolah Shirazi of Marian University’s E. S. Witchger School of Engineering, with Hung-Fu Chang from the R.B. Annis School of Engineering at the University of Indianapolis. Their pipeline starts with turning movement counts, or TMCs, derived from vision, and ends with signal schedules that adapt to how busy a city actually is across the day. The promise isn’t fancy gimmicks; it’s a practical, data-driven approach to reducing waiting times, smoothing flows, and potentially trimming the energy and emissions that come with idling cars.
At its core, the study asks a deceptively simple question: what if traffic signals could adapt to the actual way people move through an intersection, not just to a historical average? The researchers built a four-part system that starts by watching intersections with cameras, translates those observations into turning movement counts, feeds those counts into a city-scale simulator, and then tests three signaling strategies—static, dynamic, and a hybrid blend. The testbed is six Las Vegas intersections, chosen because publicly available cameras offer a realistic, day-in-the-life snapshot of real traffic. The result is not a single magic formula but a framework that shows when and where a hybrid approach—one that stays static in calmer moments but shifts to dynamic timing during peaks—can outperform either method alone. In a world of smart cities, this is a pragmatic blueprint for signals that learn to respond to the city’s pulse instead of forcing the city to fit a rigid plan.
Turning Data into Light: How TMC Feeds Signal Timings
Turn-by-turn decision-making in traffic control has long leaned on the kind of data that’s easy to collect but not always easy to act on. Traditional fixed-timing (static) signals are simple and predictable, but they can feel like a drumbeat that never changes, even when the traffic orchestra shifts. The authors’ approach foregrounds turning movement counts, the traffic volumes moving in and out of each intersection leg in specific directions (left, right, straight). They didn’t just rely on raw numbers; they built a vision-based system that identifies and follows vehicles through time. The detection uses YOLOv5—an algorithm celebrated for its speed and accuracy in real-time object recognition—while tracking relies on a Discriminative Correlation Filter approach to keep tabs on each vehicle across frames. When a vehicle exits the camera’s frame or is briefly occluded, the system gracefully handles gaps so it can still reconstruct reliable trajectories. The end product is a robust estimate of how many cars want to go West, North, East, or South at any given moment, a dataset that is otherwise painstaking to assemble by hand and easy to misread if you only count what you see from a single frame.
Crucially, the authors weathered the practical challenge that many urban cameras can only see part of an intersection at a time. They negotiated this by focusing on turning movement counts inferred from trajectories using a longest common subsequence approach, a method that tolerates disturbances and gaps in the data. The vision stage generates a turning movement count dataset for each minute of a four-hour window and exports it into the SUMO traffic simulator as a real-world-like input. To introduce realism beyond a single hour, they built a traffic generator module that can simulate a bimodal daily pattern—morning and late-afternoon peaks—by weighting flows from four zones (West, North, East, South) differently over the day. That matters: in real cities, some corridors are consistently busier than others, and the way those busy lanes interact with others can radically shift which signal strategy works best.
Once the TMC data lands in SUMO, the authors explore three signaling strategies. The static method slices each cycle into four phases with fixed green times for each phase; the dynamic method tailors green times within a cycle to the observed turning counts, effectively privileging the busiest movements in real time; and the hybrid method tries to harness the best of both worlds—using static timing during off-peak hours and switching to dynamic adjustments when congestion rises. They even layered in a reinforcement-learning approach as a point of comparison, implemented as a Deep Q-Network that allocates green time based on normalized, four-direction traffic volumes. The result is a spectrum of strategies that reveal when a city would benefit from staying on rails and when it should bend to the moment-by-moment traffic reality. The six Las Vegas intersections serve as a microcosm for broader urban networks: some are naturally more accommodating, others feel the pinch of limited lanes and asymmetric inflows, and the study’s findings hinge on those differences as much as on overall volume.
Static, Dynamic, and the Hybrid Dance
The static method is the simplest cousin of signal control. It prescribes green times for four phases across the entire simulation, based on the peak demand in non-conflicting movements. The advantage is obvious: predictable, easy to manage, and robust when traffic is evenly spread. But it’s a brittle strategy when conditions tilt—like a construction zone altering flow or a crash redirecting cars away from a preferred direction. Enter the dynamic method, which recalibrates green times by looking at the current turning movement counts across the cycle. The core idea is straightforward: if right- and left-turning volumes outpace other movements, the system extends green time for those directions to whittlequeuing and reduce delay. Yet the dynamic approach isn’t a universal fix. The researchers discovered that for intersections with lower ratios of turning movement counts to the number of lanes, pure dynamic control can underperform compared to a well-chosen static plan. In other words, there are kindling moments when too much real-time tinkering can destabilize flow rather than stabilize it.
That tension is precisely what motivates the hybrid method. It isn’t a single algorithm wearing two faces; it’s a carefully choreographed switch between static and dynamic modes, guided by the observed traffic pattern. In practical terms, the hybrid method uses a static configuration during off-peak hours and shifts to a dynamic configuration in peak times, with a mesh of the two tuned to the four-zone traffic pattern. The results show promise: when traffic is bimodal—two distinct peaks in a day—the hybrid approach tended to outperform either static or dynamic methods on several key intersections, especially where West-East and North-South flows dominate. In effect, the city’s daily pulse is not uniform; it surges and relaxes, and the hybrid method is designed to ride those ebbs and flows without flipping the switch too often or too late.
Beyond the three concrete strategies, the study also peeks at a more futuristic line of thought: could reinforcement learning, given enough high-quality, live data, push signal control toward even smarter decisions that balance throughput and fairness? The Deep Q-Network they tested showed competitive performance, particularly at the busier intersections, but it also underscored a practical caution: ML-based controllers can require substantial data and careful design to avoid overfitting to a particular day or to a narrow set of patterns. In the authors’ hands, RL remains a potent, aspirational complement rather than a guaranteed replacement for rule-based timing. The upshot is a nuanced takeaway: hybrid, data-driven strategies hold the most promise when they’re married to a robust, data-generating backbone that can reflect the city’s true rhythms—not just the memory of yesterday’s traffic.
What a City Might Earn from Learning to Adapt
The six intersections studied sit in Las Vegas, but the methodological spine of the work travels far beyond a single corner of the map. The framework is modular by design: an intersection design module, a traffic generator, a vision-based tracking system, and a signal-design module that exports configurations compatible with SUMO, a popular traffic simulator. The value isn’t just the numbers on a chart; it’s a blueprint for teams across universities and city halls to test signal strategies in a controlled, repeatable way before touching real-world intersections. In practice, what the study demonstrates is how TMC data—when captured, interpreted, and fed into a smarter simulation loop—can reveal conditions under which one signaling approach outperforms another. The key is recognizing when a city’s traffic pattern resembles a bimodal daily rhythm rather than a flat, even flow, and then letting the system adapt accordingly.
The authors kicked off their exploration with a vision-based data pipeline. Cameras, AI-based detection, and tracking yield turning movement counts with a level of granularity that’s hard to achieve with traditional loop detectors or manual counts. They didn’t stop there: by simulating four hours of traffic with zone-based distributions, they could probe how traffic patterns across West, North, East, and South zones influenced the best choice of signal strategy. The numbers matter because they point toward a practical criterion for cities: the choice of cycle length (90 seconds versus 120 seconds, for example) and the scheduling method are not one-size-fits-all; they should reflect the intersection’s geometry, lane counts, and the way traffic tends to arrive from each direction. This is where the study’s real-world relevance begins to shine: it doesn’t pretend to solve city-scale traffic in one go; it gives urban planners a tested set of tools to compare strategies in a data-driven way before committing to costly changes on the ground.
Looking ahead, the paper’s implications extend into the broader ambitions of smart cities. If cameras and ML-derived TMC data can be reliably translated into signal timings, we could see more responsive corridors that adapt not just to the moment but to the day’s arc. Dynamic and hybrid strategies might be used in concert with other smart-city levers—signal priority for buses, pedestrian safety offsets, or adaptive crosswalk timing—to reduce overall delay, improve travel times, and trim fuel consumption and emissions from idling. The road to scale remains noisy and complex—cities are mosaics of neighborhoods with different patterns, and signals interact across intersections in ways that can amplify or dampen flow. Yet the study offers a concrete path: build repeatable, camera-based data pipelines; test a spectrum of control strategies in sim-like environments; and let the daily rhythm of traffic guide the signal design rather than a single historical snapshot.
Finally, it’s worth naming the human element. The study embodies a collaboration between Marian University and the University of Indianapolis, reflecting a broader trend in which engineering, computer vision, and transport modeling converge to tackle everyday urban problems. The work is a reminder that improving city life often starts with watching the street with new eyes, learning the language of traffic movement, and translating that language into light. When a city’s lights learn to respond to the flow of its people, the journey through the urban maze can feel a touch more humane—less stop-and-start, more steady, predictable, and fair across the many neighborhoods that share the road.
In sum, Shirazi and Chang’s team aren’t proposing to replace humans with machines at the traffic light; they’re offering a way for signals to behave more like well-informed partners in the commute. The hybrid approach, tested across six real intersections with a plausible daily rhythm, provides a blueprint for practical, incremental improvements. If cities adopt and adapt this approach, we might not see a single silver bullet for traffic jams, but we could see a set of intelligent, collaborative signals that help every driver, cyclist, and pedestrian glide through town with a little less friction—and a lot more confidence in the next green light.
