An adaptive convolutional analytics architecture for closed-loop optimization in smart cities

Urban traffic corridors embody non-linear, non-stationary, and stochastic flow characteristics, governed by heterogeneous vehicular compositions, environmental perturbations, and cognitive variability in driver behavior. The resulting spatio-temporal uncertainty significantly impairs the responsiveness and scalability of conventional pre-timed or heuristic rule-based traffic control strategies, which lack adaptive inference capabilities for real-time flow modulation. To address these constraints, this study proposes a self-adaptive Intelligent Traffic Control System (TCS) that tightly couples deep convolutional vision analytics with closed-loop signal optimization. The system integrates a fine-tuned YOLOv11 detection backbone augmented with hierarchical feature fusion and attention-guided spatial refinement, enabling high-fidelity multi-scale localization and semantic classification under severe illumination fluctuations, occlusion density, and congestion turbulence. Empirical assessments performed on heterogeneous, real-world urban datasets reveal a mean detection performance of 88.45%, surpassing prior YOLO variants in precision–recall stability, temporal inference latency, and computational cost efficiency. The proposed framework leverages dynamic feedback-regulated phase modulation to estimate vehicular density ρ(t), optimize green-phase duration τ g , and maximize intersection throughput η (ρ, τ g ), thereby suppressing idle delays and mitigating congestion propagation. This research introduces context-aware, scalable, and resource-constrained traffic governance architecture, demonstrating transformative potential for next-generation smart city ecosystems by bridging perceptual intelligence with autonomous decision optimization.