Intelligent Traffic Light System Using Convolutional Neural Network to Optimize Vehicle Flow in Indonesia

Abstract

Traffic congestion remains a persistent challenge in major Indonesian cities due to rapidly increasing vehicle density and the continued reliance on fixed-time traffic signal scheduling. This study aims to develop an intelligent traffic light system that dynamically adjusts sequencing based on real-time road conditions. A dataset of 902 CCTV images was collected from intersections in six large cities across Indonesia, including Jakarta, Bandung, Surabaya, Medan, Bali, and Samarinda, and labeled into three density categories (low, medium, high). To classify vehicle density, two Convolutional Neural Network (CNN) architectures were designed and trained, incorporating preprocessing techniques such as image resizing, color inversion for illumination normalization, and data augmentation to enhance generalization. The performance of the CNNs was compared against a fuzzy logic model and a YOLOv8-based detection pipeline. Evaluation using stratified 10-fold cross-validation showed that the second CNN architecture achieved the best performance with an accuracy of 81%, precision of 0.87, recall of 0.83, and F1-score of 0.849, outperforming both baselines. Ablation studies further demonstrated that batch normalization, dropout, and data augmentation significantly reduced overfitting and improved robustness across varying light conditions. These findings indicate that a lightweight, global-context CNN can provide reliable density classification and serve as the decision engine for adaptive traffic light control. Future work will expand dataset diversity, test cross-city generalization, and explore real-time deployment in collaboration with transportation authorities to support smart city development in Indonesia.