OverFlowLight cuts urban gridlock 60.4% using real-time signal control
AI prevents cascading traffic jams by sensing overflow and inserting dedicated signal phases
OverFlowLight is a novel framework from researchers at multiple Chinese institutions that tackles urban traffic gridlock at its root cause: queue overflow. Instead of just optimizing throughput like traditional traffic signal control (TSC) algorithms, OverFlowLight uses multi-modal sensing from cameras and radars to detect when vehicle queues are about to exceed intersection capacity. Upon detection, it dynamically generates and inserts dedicated 'overflow phases' into the signal cycle to clear blocking queues before they cascade into gridlock.
The system uses a hybrid control design: a fast rule-based module for immediate overflow intervention paired with a longer-horizon reinforcement learning (RL) back end for ongoing efficiency. Deployed across 43 intersections in three major cities, OverFlowLight demonstrated seamless integration with existing RL-based TSC agents. Empirical results show a 60.4% reduction in overflow incidents and an 18.2% increase in network throughput versus deployed baselines, while also cutting the need for manual expert tuning. The project's code, datasets, and demonstration videos are publicly available.
- Uses cameras and radars to detect real-time queue overflow conditions.
- Hybrid control combines rule-based overflow phases with RL for long-term throughput.
- Deployed across 43 intersections in 3 cities; reduces incidents by 60.4% and boosts throughput 18.2%.
Why It Matters
First scalable data-driven framework actively preventing gridlock, making urban traffic more resilient and efficient.