Can adaptive traffic lights cut queues without starving side roads? A simulation exposes the trade-off
A peer-reviewed Indian study reduced average waiting and raised throughput when traffic was highly uneven—but minor approaches waited 2.5–3.5 times longer. Every control result came from simulation, not a live junction.
The short version
- What happened
- In ten-seed SUMO simulations with asymmetric demand, the controller cut mean waiting to 51.6 seconds, compared with 55.4 for fixed-time control and 66.7 for Webster's method; it served 413.9 vehicles per seed versus 311.7 and 278.4.
- Who it affects
- A controller can reduce the average while making some journeys worse. Road authorities would need public fairness rules, per-approach monitoring and safeguards for pedestrians, cyclists, buses and emergency traffic before using this design.
- What remains unproven
- Traffic control evaluated only in simulation
Keep in mind
Traffic control evaluated only in simulation
Ten stochastic seeds
Single simulated junction design
No pedestrians or cyclists
No emergency-vehicle or transit priority
No connected-junction network
No field weather or occlusion test
Plate workflow lacks privacy and redress evaluation
Early accepted article subject to production edits
Source analysis · 1 linked source. Primary analysed source date: 31 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The controller helped when one approach dominated
A peer-reviewed study from PSG College of Technology tested an adaptive signal that uses camera detections to estimate occupancy-weighted lane demand, then gives priority to the busiest approach while enforcing a maximum-wait safeguard. In Eclipse SUMO 1.27, researchers compared it with fixed-time control and Webster's traffic-engineering method under low, moderate, congested and deliberately asymmetric demand. Ten stochastic seeds were used for each setting.
The strongest result came in the condition the system was designed for: uneven arrivals. Mean waiting was 51.6 seconds, against 55.4 seconds for fixed-time and 66.7 seconds for Webster. The adaptive controller served an average 413.9 vehicles per seed, compared with 311.7 and 278.4, and cut delay on the dominant approach by 59.1% relative to fixed-time control. Under symmetric demand, however, it did not improve waiting time and recorded worse fairness than fixed timing.[1]
Average improvement concealed a side-road cost
The paper's useful contribution is that it does not present the average as a universal win. Under skewed demand, the two lightly used approaches waited 2.5–3.5 times longer. The anti-starvation constraint activated in roughly 80% of cycles, compared with 2.5–3.5% under symmetric demand, keeping those waits just below the authors' limit rather than eliminating them. That is an operational value choice: more vehicles move overall because smaller streams accept longer delays.
Camera sensing was evaluated separately. A YOLOv26 vehicle detector reported 90.92% mAP@50 and a plate detector 97.21%. A CRNN transcribed plates with 96.61% accuracy on a 1,020-image test split, ahead of three comparison architectures; the frame pipeline ran at 7.6–11.9 milliseconds. The number-plate path was proposed for automated insurance verification, but it was asynchronous and did not cause the signal decision. The paper reports recognition accuracy, not false accusation, appeal, retention, access-control or legal-compliance outcomes.[1]
The distinction that matters
What changed under asymmetric simulated demand
The controller improved network averages by prioritising the busiest approach, while a safeguard bounded—but did not remove—the cost to minor approaches.
Mean waiting
Adaptive vs fixed-time
51.6s vs 55.4s
Throughput
Vehicles served per seed
413.9 vs 311.7
Minor approaches
Delay under skew
2.5–3.5× longer
Live roads
Field deployment
Not tested
A live trial would need to test people, not only vehicles
For drivers on a persistently overloaded approach, responsive timing could shorten queues and reduce the chance of being stranded through repeated cycles. But a person entering from a quieter street could wait substantially longer. Pedestrians, cyclists, buses, emergency vehicles, junction spillback and networks of connected intersections were not part of the reported comparison. Nor does ten-seed simulation establish how camera errors, weather, occlusion or unusual behaviour affect a roadside deployment.
A credible next step would preregister a controlled field trial across different junction types, publish per-approach distributions rather than only means, and set public limits for maximum wait and protected road users. It should compare similar time periods with ordinary control, report interventions and near misses, and measure whether gains persist outside peak skew, across school hours, weekends and special events. It should also separate traffic sensing from number-plate processing and evaluate necessity, proportionality, security and redress. Until then, the study shows a quantified simulation trade-off—not proof that an AI signal will improve a real city's traffic fairly.[1]