<![CDATA[Rapid urban motorization has intensified congestion at signalized intersections, where conventional fixed-time control fails to accommodate fluctuating traffic demand. This study proposes an interpretable, real-time adaptive traffic signal system that integrates deep learning–based perception with fuzzy logic decision-making. Unlike prior works that treat detection and control as separate components, this research establishes an end-to-end perception-to-decision pipeline linking YOLOv3-based vehicle detection to a Mamdani fuzzy inference controller. Traffic videos are processed frame by frame to detect and count vehicles, from which lane-level parameters—vehicle count, queue length, and density—are extracted as fuzzy inputs. The controller adaptively determines green-phase durations according to real-time traffic states. Experiments using 300 real-world video frames under varying congestion conditions achieved precision and recall rates of 0.91 and 0.88, respectively, confirming YOLOv3’s suitability for urban traffic environments. The adaptive system produced dynamic green times ranging from 20 to 52 seconds, reducing average green duration by approximately 29% relative to fixed-time control while maintaining effective queue clearance. These findings demonstrate that the proposed integration achieves both computational efficiency and interpretability, offering a practical alternative to opaque deep reinforcement learning–based controllers. The study contributes to the growing discourse on explainable AI in transportation by operationalizing a transparent, deployable framework that links vision-based sensing to adaptive signal control, enhancing responsiveness and scalability for next-generation intelligent traffic management systems.]]>
