Prediction is Better than Detection: Traffic Congestion Control using Drones
This preprint introduces a Mesa-based multi-agent simulation where UAVs patrol road junctions to detect and predict traffic jams and trigger adaptive signal control. Across sweeps of fleet size, traffic level, and network size, detection and prediction rates plateau once drone count roughly matches junction count; prediction-triggered signal changes cut jam duration by up to 8 ticks versus up to 5 ticks for detection-triggered changes, roughly doubling benefit. It suggests onboard prediction rather than larger fleets is the higher-value path for aerial congestion management.
How does closed-loop traffic-signal control performance scale with drone fleet size, and is pre-emptive jam prediction more valuable than reactive jam detection when UAV sensing drives signal adaptation?
Existing UAV traffic monitoring is largely open-loop observation, and prior UAV-driven signal control reacts only after congestion has fully formed; fixed ground sensors are costly to deploy densely. It is unclear how fleet size translates into downstream control benefit and whether prediction adds material value beyond detection.
Prior work includes UAV and ground-sensor fusion for flow forecasting, the pNEUMA large-scale aerial sensing trial, and AVARS, a UAV-camera-fed deep reinforcement learning controller that restores already-congested networks. These treat sensing separately or act reactively, without characterizing fleet-size scaling or pre-emptive prediction for signal control.
A new simulation couples Nagel-Schreckenberg cellular-automaton vehicle dynamics, signalized four-way junctions, stochastic obstacles, and round-robin UAV patrols with hover and processing delays. Drones apply a three-variable jam detector (density, speed, queue) and a relaxed density-gradient predictor; reports trigger adaptive green-time extensions and pre-emptive shortening at a ground station. The authors sweep 1–10 drones, low/medium/high vehicle spawn rates, and 2/6/10-junction networks.
In 20 seeded runs of 2000 ticks on a 120x120-cell grid, detection rate improves with fleet size and plateaus near fleet size equal to junction count, but never reaches 100% even for 2 junctions because of 10–40 tick processing latency. At 10 junctions, the medium-vs-high traffic detection-delay gap falls from about 25 ticks with one drone to about 9 ticks with two and about 2.5 ticks with ten. Maximum prediction rate plateaus around 55% regardless of fleet size. In the 6-junction network, adapting signals after detection reduced jam duration by up to about 5 ticks at high traffic; pre-emptive adaptation after prediction reduced it by up to about 8 ticks, roughly double, while low traffic showed no measurable benefit.
Simulation only with a simplified 120x120-cell urban grid; no real-world UAV, sensor, or communication trials. The threshold predictor plateaus at about 55% and authors state the precursor signal, not fleet size, is the bottleneck. Round-robin patrol is deliberately simple and not optimal; signal, disturbance, and traffic parameters are not tuned to real-world data. The road topology is limited to signalized four-way junctions.
Could inform smart-city traffic management, UAV-based monitoring services, and adaptive signal control vendors. Before deployment, the predictive benefit must be validated on real UAVs with onboard perception, BVLOS/regulatory approval, and reliable low-latency links to signal controllers. Near-term (1–3 years) pilot integrations are plausible; citywide deployment is 3+ years.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.