Comparison of CNN Models Using EfficientNetB0, MobileNetV2, and ResNet50 for Traffic Density with Transfer Learning

Traffic congestion in urban areas poses a significant and widespread challenge, stemming from the essential role of modern transportation in daily human activities. To address this issue, artificial intelligence (AI), particularly through applying convolutional neural networks (CNN), offers a promising solution for developing automated, accurate, and efficient traffic density classification systems. However, the performance of such systems is critically dependent on the selection of optimal model architecture. This study comprehensively analyzes three leading pre-trained CNN models: EfficientNetB0, MobileNetV2, and ResNet50. Utilizing a transfer learning approach, the models were trained over 20 epochs to classify traffic density into five categories: Empty, Low, Medium, High, and Traffic Jam. The research methodology was based on the public Traffic Density Singapore dataset. To enhance model robustness and address class imbalances, the initial dataset of 4,038 images was expanded to 6,850 images through data augmentation techniques. All images were subsequently resized to a uniform size of 224x224 pixels. The evaluation results conclusively demonstrate that the ResNet50 architecture delivered superior performance, achieving a validation accuracy of approximately 85%. Furthermore, ResNet50 consistently yielded higher precision, recall, and f1-score values across most classes. For comparison, EfficientNetB0 and MobileNetV2 achieved 81% and 79% validation accuracies, respectively. This study concludes that ResNet50 is the optimal architecture for this classification task, and these findings establish a foundation for developing real-world, intelligent traffic monitoring systems.