An Original Deep Learning Architecture for Autonomous Real-Time Traffic Management in Nigerian Urban Intersections

Traffic congestion at urban intersections in Nigerian cities causes significant economic losses and road safety risks, and the adaptive traffic signal control systems required to address this problem must operate under non-standard traffic composition, unpredictable pedestrian behaviour, and unreliable sensing infrastructure. This study develops an original deep learning architecture for autonomous real-time traffic management at Nigerian urban intersections. The architecture, designated NigeriaTrafficNet, introduces four original contributions: a multi-object detection and tracking model (fine-tuned YOLOv8 with a custom Nigerian traffic class taxonomy distinguishing motorcycles, tricycles, minibuses, articulated lorries, and pedestrian modes that constitute the distinctive Nigerian traffic mix) achieving 91.2 mAP on a new Nigerian traffic dataset of 28,000 annotated intersection frames from Lagos, Kano, and Abuja; a graph convolutional network-based intersection state encoder that models the spatial relationships between approach lanes and turning movements as a dynamic graph; a deep Q-network reinforcement learning agent that learns adaptive signal timing policies from the intersection state encoder, trained in a SUMO traffic simulation environment calibrated with Nigerian traffic flow parameters; and a fault-tolerant operation protocol enabling graceful degradation to fixed-time control when camera or communication hardware fails. Simulation evaluation shows a 29 percent reduction in average vehicle delay compared to optimised fixed-time signals under Nigerian traffic conditions. Hardware deployment on a single NVIDIA Jetson AGX Orin module demonstrates real-time operation at 28 frames per second.

Keywords: deep reinforcement learning, traffic management, urban intersections, Nigeria, computer vision