Adaptive Traffic Signals

2026

Abstract

A vision and reinforcement-learning pipeline that estimates traffic density, forecasts flow, and controls signals.

View source

1   Background and motivation

A fixed signal schedule cannot respond directly to changing queues, while a simple count per lane loses information about where vehicles are accumulating. This project explores a common spatial representation for traffic perception, forecasting, and signal control. Density maps preserve both an approximate count and its location, giving the later stages more information than a single scalar measurement.

2   Density estimation and forecasting

The first stage uses a grouped-convolution residual CNN to transform an intersection image into a density map. In the documented implementation, a 72 × 72 input becomes an 18 × 18 map whose cell values sum to an estimated vehicle count. Training uses annotated Jaipur CCTV frames, with marked vehicle centres converted into target density maps.

The second stage uses ST3DNet models for 15-, 30-, and 60-minute forecasts. One stream processes recent observations and another processes observations from comparable times in preceding weeks. Their outputs are combined to predict a future map. Using the same spatial format across horizons keeps the forecasting interface consistent with the estimator, even though forecasting quality depends on the temporal data used for training.

Nested control hierarchy over a simulated road network Camera frames become density maps, which are forecast forward. The control policy nests three levels: a lane encoder inside junction-level attention, inside a graph transformer that propagates context across all thirty junctions before the actor emits a phase and a green duration. Camera frames SUMO, projected per approach Density CNN density regression, edge-sized ST3DNet forecast recent + prior-week streams Network · graph transformer all 30 junctions, 8-step temporal window Junction · set attention aggregates every approach Lane · patch CNN cross-approach attention Actor — phase selection and green duration frames density map 15 / 30 / 60 min MAPPO simulated network, not deployed hardware
Figure 1.The control policy is nested rather than sequential: a lane encoder sits inside junction-level attention, which sits inside a graph transformer spanning all thirty junctions. A signal decision is therefore made with network context, not from one junction’s own queue. Results are from the simulated network.

3   Network-level signal control

The control stage runs in SUMO and uses a hierarchical network to combine information from incoming approaches, junctions, and neighbouring signals. Junction adjacency is read from the simulation network, allowing the representation to follow the supplied road graph. MAPPO supplies the multi-agent training procedure, with a centralised critic supporting coordination across intersections.

A projection component converts simulated vehicle positions into grid representations for the controller. This gives the policy a spatial input resembling the earlier stages without requiring live camera feeds during every simulation run. The resulting control experiments can be repeated under a specified simulation configuration, which is essential for comparing policies under similar traffic conditions.

4   Experimental scope

The stages have different evidence behind them. The repository documents a density estimator trained on real images, forecasters trained on synthetic data, and signal-control training and evaluation in simulation. These should not be presented as an end-to-end road deployment. The main contribution is the staged experimental pipeline; validating transfer to real intersections would require further data, evaluation, and integration beyond the simulation results.