Fairness aware real time traffic signal control using vehicle detection under symmetric and asymmetric demand

Urban traffic congestion imposes measurable costs in travel time, fuel consumption and vehicular emissions. Fixed-time signal controllers, which remain the dominant form of intersection control, follow a schedule computed offline and cannot respond to instantaneous demand. Adaptive controllers address this, but are commonly reported on aggregate efficiency alone, without measuring how delay is distributed across approaches, so a controller may improve mean delay while starving a minor approach. This work presents an integrated framework combining camera-based sensing, a priority-based adaptive signal controller with an explicit anti-starvation constraint, and an asynchronous number plate recognition path for automated insurance verification, and subjects the controller to the quantitative evaluation that comparable systems generally omit. Sensing requires no in-road instrumentation. A YOLOv26 detection model classifies vehicles into four categories and produces occupancy-weighted lane densities; a second YOLOv26 model localises number plates for a recognition path that transcribes plate text using a Convolutional Recurrent Neural Network with Connectionist Temporal Classification decoding. The vehicle detector achieves an mAP@50 of 90.92% and the plate detector 97.21%. Among four recognition architectures trained on identical data the CRNN attains the highest accuracy at 96.61%, ahead of a Transformer encoder-decoder at 94.37%, a recurrent baseline at 91.84% and a convolutional baseline at 86.59%; every pairwise difference is statistically significant on the 1,020-image test split (two-proportion z-test, p < 0.05). The detection pipeline processes a frame in 7.6–11.9 ms, approximately 84–130 FPS. The controller was evaluated in Eclipse SUMO 1.27 against fixed-time and Webster’s-method baselines across low, moderate, congested and asymmetric demand, with ten-seed replication under stochastic arrivals. The results are scenario-dependent and are reported as such. Under symmetric demand the proposed controller does not improve on fixed-time control, recording higher mean waiting time and a lower Jain’s Fairness Index, because uniform loading offers no inequity for a demand-responsive policy to correct. Under asymmetric demand, the condition the controller is designed for, it records the lowest mean waiting time at 51.6 s against 55.4 s and 66.7 s for the baselines, improves throughput by 28.0% and 51.8% respectively, and serves 413.9 vehicles per seed against 311.7 and 278.4. The dominant approach’s delay falls by 59.1% against fixed-time control and its demand is fully served, where the baselines leave 98–136 vehicles per seed stranded. The explicit cost, disclosed rather than smoothed over, is that the two minor approaches wait 2.5–3.5 times longer, bounded just below the starvation threshold by the anti-starvation constraint, which activates in approximately 80% of cycles under skewed demand against 2.5–3.5% under symmetric demand. The contribution is a quantified characterisation of this exchange, together with an ablation establishing occupancy-weighted density as the component carrying the controller’s performance.