Advances in Smart Traffic Signal Control: A Comprehensive Survey

The rapid increase in vehicle numbers and the dynamic nature of traffic patterns have made road traffic management a significant global challenge. Conventional traffic signal systems are unable to adapt to real-time conditions, resulting in unnecessary delays, increased congestion, and higher fuel consumption. In the present study, we review recent advancements in smart traffic signal control systems. Our study highlights the progression from supervised machine learning approaches (ATLCS, priority-based adaptive controls, genetic algorithm-driven optimization models), to deep learning methods (2D-CNN-based vehicle classification), and to reinforcement learning techniques (IntelliLight, FRAP, PressLight, AttendLight, robust DQN-based frameworks, MACOPO, MAGSAC) designed to learn traffic patterns, adapt to dynamic environments, and make data-driven decisions in real time—capabilities not feasible with conventional methods. We divide the scope into machine learning–based, reinforcement learning–based, and hybrid approaches. Our analysis was structured to compare how each model improves traffic efficiency using key performance attributes, including travel time, queue length, waiting time, fuel consumption, and intersection throughput. These models have proven effective at cutting down delays, reducing the length of queues, and enhancing overall traffic flow by adjusting to dynamic traffic conditions. For instance, IntelliLight achieved 72% reduction in travel time compared to fixed-time traffic control, along with significant decreases in average delay. These approaches perform well in different settings, from small intersections to large-scale networks, because they are effective, flexible, and scalable. Overall, our studies highlight that recent advancements in Artificial Intelligence (AI) are crucial for traffic management, with great potential to make urban transportation faster, safer, and more environmentally friendly. We also identify the limitations of these models and suggest future research directions.