Intelligent Traffic Light Optimization System Using Convolutional Neural Networks for Historic City Centers in Complex Scenarios

Tenorio-Huarancca D.O.; Lizarbe-Alarcon H.; Ayala-Bizarro R.G.; Tenorio-Palomino M.G.; Bravo-Anaya R.G.; Ircañaupa-Huamaní A.S.; León-Palacios E.; Bellido-Aedo V. · Mathematical Modelling of Engineering Problems (2025) · DOI: 10.18280/mmep.120928

Resumen

Historic urban centers present a paradigmatic challenge in modern traffic management, characterized by narrow streets originally conceived for carriage and pedestrian circulation. This infrastructural incompatibility generates critical congestion, exacerbated in developing countries where fixed-time traffic signal systems predominate, lacking adaptive capacity and generating substantial inefficiencies of temporal, energy, and fuel resources. We developed a convolutional neural network model based on a customized You Only Look Once version 8 architecture for vehicle detection and classification. The model implements advanced temporal filtering to reduce false positives, vehicle tracking for unique counting, and a comprehensive 13-stage traffic signal optimization algorithm that correlates detected vehicular density with cycle times. The system maintains operational robustness under adverse conditions, including precipitation, cloudiness, shadows cast by colonial mansions, vehicular occlusion phenomena, and luminous glare. Implementation was evaluated through video recordings from the Historic Center of Ayacucho, using strategically positioned cameras to determine vehicular density at various time periods. The model, trained for 126 epochs with Early Stopping on 3,000 images, achieves 88.7% precision, recall of 0.832/0.834 (validation/evaluation), establishing a robust solution for urban heritage contexts.

Instituciones: Universidad Nacional de San Cristóbal de Huamanga