OverFlowLight cuts urban gridlock 60% with real-time queue detection
A real-time traffic signal framework deployed in three cities uses camera-radar fusion to detect queue overflow and inject dedicated clearance phases, reducing gridlock incidents by 60.4% while boosting throughput 18.2%.

OverFlowLight, a real-time traffic signal control framework, deployed across 43 intersections in three major cities and cut overflow incidents by 60.4 percent while increasing network throughput by 18.2 percent compared to existing signal plans. The system detects when vehicle queues exceed intersection capacity—a condition that blocks upstream traffic and triggers cascading gridlocks—by fusing data from cameras and radars, then dynamically inserts dedicated overflow phases into the signal cycle to clear the blocking queues before they spread.
Queue overflow is a distinct failure mode from general congestion: it occurs when a queue physically exceeds the space between intersections, obstructing cross-traffic and creating safety hazards. Prevailing traffic signal algorithms optimize for vehicle throughput but lack mechanisms to detect or preempt overflow in real time. OverFlowLight pairs rapid rule-based intervention for immediate overflow response with longer-horizon reinforcement learning backends for efficiency. That hybrid design lets it integrate with existing RL-based traffic signal controllers without replacing them, a modularity the team validated in production. The framework substantially reduced the need for manual intervention that expert-tuned signal plans typically require during peak hours.
The camera-radar fusion generalizes across different sensor setups, and the modular design means cities can layer it onto existing RL-based controllers. The open question is how the 60 percent reduction holds in denser networks with tighter intersection spacing, and whether the approach scales beyond the 43-intersection pilot. Full metropolitan rollout will likely require integration with citywide sensor networks and coordination across signal timing domains—challenges the team has flagged but not yet addressed in the preprint.


