Efficient Deep Learning for Accurate Vehicle Classification in Urban Computing

Abstract

The rapid growth of urbanization has increased the need for reliable and automated vehicle classification systems to support intelligent transportation and urban computing applications. This study presents a comparative deep learning framework for classifying vehicle images into four categories: bicycle, bike, car, and truck. Seven widely used convolutional neural network (CNN) architectures VGG-16, VGG-19, Inception-V3, Inception-V4, ResNet-50, ResNet-101, and EfficientNet-B0 are evaluated using transfer learning and data augmentation. The dataset comprises RGB vehicle images collected from online image sources and organized into training and test sets. During training, image rescaling, shear transformation, zooming, and horizontal flipping are applied to increase training-data variability, while the test data are evaluated without geometric augmentation. Model performance is assessed using accuracy, loss, precision, recall, and F1-score. Among the evaluated architectures, EfficientNet-B0 achieves the highest reported classification accuracy of 97.06%, followed by ResNet-101 with 96.82%. EfficientNet-B0 also demonstrates strong class-level performance across all four vehicle categories, indicating its effectiveness for image-based vehicle classification. The comparative results provide insights into the suitability of different CNN architectures for vehicle classification within urban computing applications. However, the study evaluates image-level classification rather than object localization and does not explicitly assess challenging real-world conditions such as severe occlusion, nighttime illumination, or diverse camera viewpoints. Future work will therefore focus on larger and more diverse traffic datasets, computational-efficiency evaluation, real-world traffic-camera imagery, and deployment-oriented optimization for intelligent transportation systems.