A simulation-based comparative evaluation of CNN–, LSTM-, and reinforcement learning–based traffic signal control using vision-derived data

Urban signalized intersections are critical bottlenecks in metropolitan road networks, where inefficient control strategies contribute to excessive delay, congestion, and environmental impacts. Although Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) models, and Reinforcement Learning (RL) approaches have been widely studied for traffic analysis and signal control, prior research typically evaluates these paradigms in isolation using heterogeneous data sources and metrics. This fragmentation limits systematic comparison and obscures practical trade-offs. This study presents a controlled comparative evaluation of CNN–, LSTM-, and RL-based controllers for real-time traffic signal control using online vision-derived data under consistent experimental conditions. All models were evaluated at a real four-leg urban intersection (17 Shahrivar–Taleghani, Tabriz) across 12 traffic scenarios representing varying demand levels and stochastic arrival patterns, totaling 3,600 simulation episodes and more than 2.5 million control decisions. Performance was assessed using average vehicle delay, queue length, throughput, robustness metrics, and statistical significance testing. Results indicate that the RL-based controller consistently outperforms the CNN and LSTM models, reducing average delay by approximately 21–33% relative to the baselines while simultaneously increasing throughput and maintaining the lowest performance variance. The CNN-based controller demonstrated high computational efficiency but lacked anticipatory capability, whereas the LSTM-based controller improved stability yet remained constrained by its predict-then-control architecture. The RL framework achieved real-time feasibility, with 100% of decisions executed within 20 ms. The findings demonstrate that objective-aligned, decision-centric learning provides structural advantages over perception or prediction-driven paradigms. This study offers one of the first unified vision-based benchmarks across these modelling frameworks and provides evidence-based guidance for scalable adaptive traffic signal control.